diff --git a/comfy/cli_args.py b/comfy/cli_args.py
index b4f22f31..fda24543 100644
--- a/comfy/cli_args.py
+++ b/comfy/cli_args.py
@@ -54,7 +54,8 @@ fp_group.add_argument("--force-fp16", action="store_true", help="Force fp16.")
 
 fpvae_group = parser.add_mutually_exclusive_group()
 fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.")
-fpvae_group.add_argument("--bf16-vae", action="store_true", help="Run the VAE in bf16, might lower quality.")
+fpvae_group.add_argument("--fp32-vae", action="store_true", help="Run the VAE in full precision fp32.")
+fpvae_group.add_argument("--bf16-vae", action="store_true", help="Run the VAE in bf16.")
 
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")
 
diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py
index a887e51b..daaa2f2b 100644
--- a/comfy/clip_vision.py
+++ b/comfy/clip_vision.py
@@ -2,14 +2,27 @@ from transformers import CLIPVisionModelWithProjection, CLIPVisionConfig, CLIPIm
 from .utils import load_torch_file, transformers_convert
 import os
 import torch
+import contextlib
+
 import comfy.ops
+import comfy.model_patcher
+import comfy.model_management
 
 class ClipVisionModel():
     def __init__(self, json_config):
         config = CLIPVisionConfig.from_json_file(json_config)
-        with comfy.ops.use_comfy_ops():
+        self.load_device = comfy.model_management.text_encoder_device()
+        offload_device = comfy.model_management.text_encoder_offload_device()
+        self.dtype = torch.float32
+        if comfy.model_management.should_use_fp16(self.load_device, prioritize_performance=False):
+            self.dtype = torch.float16
+
+        with comfy.ops.use_comfy_ops(offload_device, self.dtype):
             with modeling_utils.no_init_weights():
                 self.model = CLIPVisionModelWithProjection(config)
+        self.model.to(self.dtype)
+
+        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
         self.processor = CLIPImageProcessor(crop_size=224,
                                             do_center_crop=True,
                                             do_convert_rgb=True,
@@ -27,7 +40,21 @@ class ClipVisionModel():
         img = torch.clip((255. * image), 0, 255).round().int()
         img = list(map(lambda a: a, img))
         inputs = self.processor(images=img, return_tensors="pt")
-        outputs = self.model(**inputs)
+        comfy.model_management.load_model_gpu(self.patcher)
+        pixel_values = inputs['pixel_values'].to(self.load_device)
+
+        if self.dtype != torch.float32:
+            precision_scope = torch.autocast
+        else:
+            precision_scope = lambda a, b: contextlib.nullcontext(a)
+
+        with precision_scope(comfy.model_management.get_autocast_device(self.load_device), torch.float32):
+            outputs = self.model(pixel_values=pixel_values)
+
+        for k in outputs:
+            t = outputs[k]
+            if t is not None:
+                outputs[k] = t.cpu()
         return outputs
 
 def convert_to_transformers(sd, prefix):
diff --git a/comfy/controlnet.py b/comfy/controlnet.py
new file mode 100644
index 00000000..490be6bb
--- /dev/null
+++ b/comfy/controlnet.py
@@ -0,0 +1,480 @@
+import torch
+import math
+import os
+import comfy.utils
+import comfy.model_management
+import comfy.model_detection
+import comfy.model_patcher
+
+import comfy.cldm.cldm
+import comfy.t2i_adapter.adapter
+
+
+def broadcast_image_to(tensor, target_batch_size, batched_number):
+    current_batch_size = tensor.shape[0]
+    #print(current_batch_size, target_batch_size)
+    if current_batch_size == 1:
+        return tensor
+
+    per_batch = target_batch_size // batched_number
+    tensor = tensor[:per_batch]
+
+    if per_batch > tensor.shape[0]:
+        tensor = torch.cat([tensor] * (per_batch // tensor.shape[0]) + [tensor[:(per_batch % tensor.shape[0])]], dim=0)
+
+    current_batch_size = tensor.shape[0]
+    if current_batch_size == target_batch_size:
+        return tensor
+    else:
+        return torch.cat([tensor] * batched_number, dim=0)
+
+class ControlBase:
+    def __init__(self, device=None):
+        self.cond_hint_original = None
+        self.cond_hint = None
+        self.strength = 1.0
+        self.timestep_percent_range = (1.0, 0.0)
+        self.timestep_range = None
+
+        if device is None:
+            device = comfy.model_management.get_torch_device()
+        self.device = device
+        self.previous_controlnet = None
+        self.global_average_pooling = False
+
+    def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(1.0, 0.0)):
+        self.cond_hint_original = cond_hint
+        self.strength = strength
+        self.timestep_percent_range = timestep_percent_range
+        return self
+
+    def pre_run(self, model, percent_to_timestep_function):
+        self.timestep_range = (percent_to_timestep_function(self.timestep_percent_range[0]), percent_to_timestep_function(self.timestep_percent_range[1]))
+        if self.previous_controlnet is not None:
+            self.previous_controlnet.pre_run(model, percent_to_timestep_function)
+
+    def set_previous_controlnet(self, controlnet):
+        self.previous_controlnet = controlnet
+        return self
+
+    def cleanup(self):
+        if self.previous_controlnet is not None:
+            self.previous_controlnet.cleanup()
+        if self.cond_hint is not None:
+            del self.cond_hint
+            self.cond_hint = None
+        self.timestep_range = None
+
+    def get_models(self):
+        out = []
+        if self.previous_controlnet is not None:
+            out += self.previous_controlnet.get_models()
+        return out
+
+    def copy_to(self, c):
+        c.cond_hint_original = self.cond_hint_original
+        c.strength = self.strength
+        c.timestep_percent_range = self.timestep_percent_range
+
+    def inference_memory_requirements(self, dtype):
+        if self.previous_controlnet is not None:
+            return self.previous_controlnet.inference_memory_requirements(dtype)
+        return 0
+
+    def control_merge(self, control_input, control_output, control_prev, output_dtype):
+        out = {'input':[], 'middle':[], 'output': []}
+
+        if control_input is not None:
+            for i in range(len(control_input)):
+                key = 'input'
+                x = control_input[i]
+                if x is not None:
+                    x *= self.strength
+                    if x.dtype != output_dtype:
+                        x = x.to(output_dtype)
+                out[key].insert(0, x)
+
+        if control_output is not None:
+            for i in range(len(control_output)):
+                if i == (len(control_output) - 1):
+                    key = 'middle'
+                    index = 0
+                else:
+                    key = 'output'
+                    index = i
+                x = control_output[i]
+                if x is not None:
+                    if self.global_average_pooling:
+                        x = torch.mean(x, dim=(2, 3), keepdim=True).repeat(1, 1, x.shape[2], x.shape[3])
+
+                    x *= self.strength
+                    if x.dtype != output_dtype:
+                        x = x.to(output_dtype)
+
+                out[key].append(x)
+        if control_prev is not None:
+            for x in ['input', 'middle', 'output']:
+                o = out[x]
+                for i in range(len(control_prev[x])):
+                    prev_val = control_prev[x][i]
+                    if i >= len(o):
+                        o.append(prev_val)
+                    elif prev_val is not None:
+                        if o[i] is None:
+                            o[i] = prev_val
+                        else:
+                            o[i] += prev_val
+        return out
+
+class ControlNet(ControlBase):
+    def __init__(self, control_model, global_average_pooling=False, device=None):
+        super().__init__(device)
+        self.control_model = control_model
+        self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device())
+        self.global_average_pooling = global_average_pooling
+
+    def get_control(self, x_noisy, t, cond, batched_number):
+        control_prev = None
+        if self.previous_controlnet is not None:
+            control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
+
+        if self.timestep_range is not None:
+            if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
+                if control_prev is not None:
+                    return control_prev
+                else:
+                    return None
+
+        output_dtype = x_noisy.dtype
+        if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]:
+            if self.cond_hint is not None:
+                del self.cond_hint
+            self.cond_hint = None
+            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * 8, x_noisy.shape[2] * 8, 'nearest-exact', "center").to(self.control_model.dtype).to(self.device)
+        if x_noisy.shape[0] != self.cond_hint.shape[0]:
+            self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
+
+
+        context = cond['c_crossattn']
+        y = cond.get('c_adm', None)
+        if y is not None:
+            y = y.to(self.control_model.dtype)
+        control = self.control_model(x=x_noisy.to(self.control_model.dtype), hint=self.cond_hint, timesteps=t, context=context.to(self.control_model.dtype), y=y)
+        return self.control_merge(None, control, control_prev, output_dtype)
+
+    def copy(self):
+        c = ControlNet(self.control_model, global_average_pooling=self.global_average_pooling)
+        self.copy_to(c)
+        return c
+
+    def get_models(self):
+        out = super().get_models()
+        out.append(self.control_model_wrapped)
+        return out
+
+class ControlLoraOps:
+    class Linear(torch.nn.Module):
+        def __init__(self, in_features: int, out_features: int, bias: bool = True,
+                    device=None, dtype=None) -> None:
+            factory_kwargs = {'device': device, 'dtype': dtype}
+            super().__init__()
+            self.in_features = in_features
+            self.out_features = out_features
+            self.weight = None
+            self.up = None
+            self.down = None
+            self.bias = None
+
+        def forward(self, input):
+            if self.up is not None:
+                return torch.nn.functional.linear(input, self.weight.to(input.device) + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), self.bias)
+            else:
+                return torch.nn.functional.linear(input, self.weight.to(input.device), self.bias)
+
+    class Conv2d(torch.nn.Module):
+        def __init__(
+            self,
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=1,
+            padding=0,
+            dilation=1,
+            groups=1,
+            bias=True,
+            padding_mode='zeros',
+            device=None,
+            dtype=None
+        ):
+            super().__init__()
+            self.in_channels = in_channels
+            self.out_channels = out_channels
+            self.kernel_size = kernel_size
+            self.stride = stride
+            self.padding = padding
+            self.dilation = dilation
+            self.transposed = False
+            self.output_padding = 0
+            self.groups = groups
+            self.padding_mode = padding_mode
+
+            self.weight = None
+            self.bias = None
+            self.up = None
+            self.down = None
+
+
+        def forward(self, input):
+            if self.up is not None:
+                return torch.nn.functional.conv2d(input, self.weight.to(input.device) + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), self.bias, self.stride, self.padding, self.dilation, self.groups)
+            else:
+                return torch.nn.functional.conv2d(input, self.weight.to(input.device), self.bias, self.stride, self.padding, self.dilation, self.groups)
+
+    def conv_nd(self, dims, *args, **kwargs):
+        if dims == 2:
+            return self.Conv2d(*args, **kwargs)
+        else:
+            raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class ControlLora(ControlNet):
+    def __init__(self, control_weights, global_average_pooling=False, device=None):
+        ControlBase.__init__(self, device)
+        self.control_weights = control_weights
+        self.global_average_pooling = global_average_pooling
+
+    def pre_run(self, model, percent_to_timestep_function):
+        super().pre_run(model, percent_to_timestep_function)
+        controlnet_config = model.model_config.unet_config.copy()
+        controlnet_config.pop("out_channels")
+        controlnet_config["hint_channels"] = self.control_weights["input_hint_block.0.weight"].shape[1]
+        controlnet_config["operations"] = ControlLoraOps()
+        self.control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)
+        dtype = model.get_dtype()
+        self.control_model.to(dtype)
+        self.control_model.to(comfy.model_management.get_torch_device())
+        diffusion_model = model.diffusion_model
+        sd = diffusion_model.state_dict()
+        cm = self.control_model.state_dict()
+
+        for k in sd:
+            weight = comfy.model_management.resolve_lowvram_weight(sd[k], diffusion_model, k)
+            try:
+                comfy.utils.set_attr(self.control_model, k, weight)
+            except:
+                pass
+
+        for k in self.control_weights:
+            if k not in {"lora_controlnet"}:
+                comfy.utils.set_attr(self.control_model, k, self.control_weights[k].to(dtype).to(comfy.model_management.get_torch_device()))
+
+    def copy(self):
+        c = ControlLora(self.control_weights, global_average_pooling=self.global_average_pooling)
+        self.copy_to(c)
+        return c
+
+    def cleanup(self):
+        del self.control_model
+        self.control_model = None
+        super().cleanup()
+
+    def get_models(self):
+        out = ControlBase.get_models(self)
+        return out
+
+    def inference_memory_requirements(self, dtype):
+        return comfy.utils.calculate_parameters(self.control_weights) * comfy.model_management.dtype_size(dtype) + ControlBase.inference_memory_requirements(self, dtype)
+
+def load_controlnet(ckpt_path, model=None):
+    controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
+    if "lora_controlnet" in controlnet_data:
+        return ControlLora(controlnet_data)
+
+    controlnet_config = None
+    if "controlnet_cond_embedding.conv_in.weight" in controlnet_data: #diffusers format
+        use_fp16 = comfy.model_management.should_use_fp16()
+        controlnet_config = comfy.model_detection.unet_config_from_diffusers_unet(controlnet_data, use_fp16)
+        diffusers_keys = comfy.utils.unet_to_diffusers(controlnet_config)
+        diffusers_keys["controlnet_mid_block.weight"] = "middle_block_out.0.weight"
+        diffusers_keys["controlnet_mid_block.bias"] = "middle_block_out.0.bias"
+
+        count = 0
+        loop = True
+        while loop:
+            suffix = [".weight", ".bias"]
+            for s in suffix:
+                k_in = "controlnet_down_blocks.{}{}".format(count, s)
+                k_out = "zero_convs.{}.0{}".format(count, s)
+                if k_in not in controlnet_data:
+                    loop = False
+                    break
+                diffusers_keys[k_in] = k_out
+            count += 1
+
+        count = 0
+        loop = True
+        while loop:
+            suffix = [".weight", ".bias"]
+            for s in suffix:
+                if count == 0:
+                    k_in = "controlnet_cond_embedding.conv_in{}".format(s)
+                else:
+                    k_in = "controlnet_cond_embedding.blocks.{}{}".format(count - 1, s)
+                k_out = "input_hint_block.{}{}".format(count * 2, s)
+                if k_in not in controlnet_data:
+                    k_in = "controlnet_cond_embedding.conv_out{}".format(s)
+                    loop = False
+                diffusers_keys[k_in] = k_out
+            count += 1
+
+        new_sd = {}
+        for k in diffusers_keys:
+            if k in controlnet_data:
+                new_sd[diffusers_keys[k]] = controlnet_data.pop(k)
+
+        leftover_keys = controlnet_data.keys()
+        if len(leftover_keys) > 0:
+            print("leftover keys:", leftover_keys)
+        controlnet_data = new_sd
+
+    pth_key = 'control_model.zero_convs.0.0.weight'
+    pth = False
+    key = 'zero_convs.0.0.weight'
+    if pth_key in controlnet_data:
+        pth = True
+        key = pth_key
+        prefix = "control_model."
+    elif key in controlnet_data:
+        prefix = ""
+    else:
+        net = load_t2i_adapter(controlnet_data)
+        if net is None:
+            print("error checkpoint does not contain controlnet or t2i adapter data", ckpt_path)
+        return net
+
+    if controlnet_config is None:
+        use_fp16 = comfy.model_management.should_use_fp16()
+        controlnet_config = comfy.model_detection.model_config_from_unet(controlnet_data, prefix, use_fp16).unet_config
+    controlnet_config.pop("out_channels")
+    controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
+    control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)
+
+    if pth:
+        if 'difference' in controlnet_data:
+            if model is not None:
+                comfy.model_management.load_models_gpu([model])
+                model_sd = model.model_state_dict()
+                for x in controlnet_data:
+                    c_m = "control_model."
+                    if x.startswith(c_m):
+                        sd_key = "diffusion_model.{}".format(x[len(c_m):])
+                        if sd_key in model_sd:
+                            cd = controlnet_data[x]
+                            cd += model_sd[sd_key].type(cd.dtype).to(cd.device)
+            else:
+                print("WARNING: Loaded a diff controlnet without a model. It will very likely not work.")
+
+        class WeightsLoader(torch.nn.Module):
+            pass
+        w = WeightsLoader()
+        w.control_model = control_model
+        missing, unexpected = w.load_state_dict(controlnet_data, strict=False)
+    else:
+        missing, unexpected = control_model.load_state_dict(controlnet_data, strict=False)
+    print(missing, unexpected)
+
+    if use_fp16:
+        control_model = control_model.half()
+
+    global_average_pooling = False
+    filename = os.path.splitext(ckpt_path)[0]
+    if filename.endswith("_shuffle") or filename.endswith("_shuffle_fp16"): #TODO: smarter way of enabling global_average_pooling
+        global_average_pooling = True
+
+    control = ControlNet(control_model, global_average_pooling=global_average_pooling)
+    return control
+
+class T2IAdapter(ControlBase):
+    def __init__(self, t2i_model, channels_in, device=None):
+        super().__init__(device)
+        self.t2i_model = t2i_model
+        self.channels_in = channels_in
+        self.control_input = None
+
+    def scale_image_to(self, width, height):
+        unshuffle_amount = self.t2i_model.unshuffle_amount
+        width = math.ceil(width / unshuffle_amount) * unshuffle_amount
+        height = math.ceil(height / unshuffle_amount) * unshuffle_amount
+        return width, height
+
+    def get_control(self, x_noisy, t, cond, batched_number):
+        control_prev = None
+        if self.previous_controlnet is not None:
+            control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
+
+        if self.timestep_range is not None:
+            if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
+                if control_prev is not None:
+                    return control_prev
+                else:
+                    return {}
+
+        if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]:
+            if self.cond_hint is not None:
+                del self.cond_hint
+            self.control_input = None
+            self.cond_hint = None
+            width, height = self.scale_image_to(x_noisy.shape[3] * 8, x_noisy.shape[2] * 8)
+            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, width, height, 'nearest-exact', "center").float().to(self.device)
+            if self.channels_in == 1 and self.cond_hint.shape[1] > 1:
+                self.cond_hint = torch.mean(self.cond_hint, 1, keepdim=True)
+        if x_noisy.shape[0] != self.cond_hint.shape[0]:
+            self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
+        if self.control_input is None:
+            self.t2i_model.to(x_noisy.dtype)
+            self.t2i_model.to(self.device)
+            self.control_input = self.t2i_model(self.cond_hint.to(x_noisy.dtype))
+            self.t2i_model.cpu()
+
+        control_input = list(map(lambda a: None if a is None else a.clone(), self.control_input))
+        mid = None
+        if self.t2i_model.xl == True:
+            mid = control_input[-1:]
+            control_input = control_input[:-1]
+        return self.control_merge(control_input, mid, control_prev, x_noisy.dtype)
+
+    def copy(self):
+        c = T2IAdapter(self.t2i_model, self.channels_in)
+        self.copy_to(c)
+        return c
+
+def load_t2i_adapter(t2i_data):
+    keys = t2i_data.keys()
+    if 'adapter' in keys:
+        t2i_data = t2i_data['adapter']
+        keys = t2i_data.keys()
+    if "body.0.in_conv.weight" in keys:
+        cin = t2i_data['body.0.in_conv.weight'].shape[1]
+        model_ad = comfy.t2i_adapter.adapter.Adapter_light(cin=cin, channels=[320, 640, 1280, 1280], nums_rb=4)
+    elif 'conv_in.weight' in keys:
+        cin = t2i_data['conv_in.weight'].shape[1]
+        channel = t2i_data['conv_in.weight'].shape[0]
+        ksize = t2i_data['body.0.block2.weight'].shape[2]
+        use_conv = False
+        down_opts = list(filter(lambda a: a.endswith("down_opt.op.weight"), keys))
+        if len(down_opts) > 0:
+            use_conv = True
+        xl = False
+        if cin == 256 or cin == 768:
+            xl = True
+        model_ad = comfy.t2i_adapter.adapter.Adapter(cin=cin, channels=[channel, channel*2, channel*4, channel*4][:4], nums_rb=2, ksize=ksize, sk=True, use_conv=use_conv, xl=xl)
+    else:
+        return None
+    missing, unexpected = model_ad.load_state_dict(t2i_data)
+    if len(missing) > 0:
+        print("t2i missing", missing)
+
+    if len(unexpected) > 0:
+        print("t2i unexpected", unexpected)
+
+    return T2IAdapter(model_ad, model_ad.input_channels)
diff --git a/comfy/diffusers_load.py b/comfy/diffusers_load.py
index 11d94c34..a52e0102 100644
--- a/comfy/diffusers_load.py
+++ b/comfy/diffusers_load.py
@@ -1,87 +1,36 @@
 import json
 import os
-import yaml
 
-import folder_paths
-from comfy.sd import load_checkpoint
-import os.path as osp
-import re
-import torch
-from safetensors.torch import load_file, save_file
-from . import diffusers_convert
+import comfy.sd
 
+def first_file(path, filenames):
+    for f in filenames:
+        p = os.path.join(path, f)
+        if os.path.exists(p):
+            return p
+    return None
 
-def load_diffusers(model_path, fp16=True, output_vae=True, output_clip=True, embedding_directory=None):
-    diffusers_unet_conf = json.load(open(osp.join(model_path, "unet/config.json")))
-    diffusers_scheduler_conf = json.load(open(osp.join(model_path, "scheduler/scheduler_config.json")))
+def load_diffusers(model_path, output_vae=True, output_clip=True, embedding_directory=None):
+    diffusion_model_names = ["diffusion_pytorch_model.fp16.safetensors", "diffusion_pytorch_model.safetensors", "diffusion_pytorch_model.fp16.bin", "diffusion_pytorch_model.bin"]
+    unet_path = first_file(os.path.join(model_path, "unet"), diffusion_model_names)
+    vae_path = first_file(os.path.join(model_path, "vae"), diffusion_model_names)
 
-    # magic
-    v2 = diffusers_unet_conf["sample_size"] == 96
-    if 'prediction_type' in diffusers_scheduler_conf:
-        v_pred = diffusers_scheduler_conf['prediction_type'] == 'v_prediction'
+    text_encoder_model_names = ["model.fp16.safetensors", "model.safetensors", "pytorch_model.fp16.bin", "pytorch_model.bin"]
+    text_encoder1_path = first_file(os.path.join(model_path, "text_encoder"), text_encoder_model_names)
+    text_encoder2_path = first_file(os.path.join(model_path, "text_encoder_2"), text_encoder_model_names)
 
-    if v2:
-        if v_pred:
-            config_path = folder_paths.get_full_path("configs", 'v2-inference-v.yaml')
-        else:
-            config_path = folder_paths.get_full_path("configs", 'v2-inference.yaml')
-    else:
-        config_path = folder_paths.get_full_path("configs", 'v1-inference.yaml')
+    text_encoder_paths = [text_encoder1_path]
+    if text_encoder2_path is not None:
+        text_encoder_paths.append(text_encoder2_path)
 
-    with open(config_path, 'r') as stream:
-        config = yaml.safe_load(stream)
+    unet = comfy.sd.load_unet(unet_path)
 
-    model_config_params = config['model']['params']
-    clip_config = model_config_params['cond_stage_config']
-    scale_factor = model_config_params['scale_factor']
-    vae_config = model_config_params['first_stage_config']
-    vae_config['scale_factor'] = scale_factor
-    model_config_params["unet_config"]["params"]["use_fp16"] = fp16
+    clip = None
+    if output_clip:
+        clip = comfy.sd.load_clip(text_encoder_paths, embedding_directory=embedding_directory)
 
-    unet_path = osp.join(model_path, "unet", "diffusion_pytorch_model.safetensors")
-    vae_path = osp.join(model_path, "vae", "diffusion_pytorch_model.safetensors")
-    text_enc_path = osp.join(model_path, "text_encoder", "model.safetensors")
+    vae = None
+    if output_vae:
+        vae = comfy.sd.VAE(ckpt_path=vae_path)
 
-    # Load models from safetensors if it exists, if it doesn't pytorch
-    if osp.exists(unet_path):
-        unet_state_dict = load_file(unet_path, device="cpu")
-    else:
-        unet_path = osp.join(model_path, "unet", "diffusion_pytorch_model.bin")
-        unet_state_dict = torch.load(unet_path, map_location="cpu")
-
-    if osp.exists(vae_path):
-        vae_state_dict = load_file(vae_path, device="cpu")
-    else:
-        vae_path = osp.join(model_path, "vae", "diffusion_pytorch_model.bin")
-        vae_state_dict = torch.load(vae_path, map_location="cpu")
-
-    if osp.exists(text_enc_path):
-        text_enc_dict = load_file(text_enc_path, device="cpu")
-    else:
-        text_enc_path = osp.join(model_path, "text_encoder", "pytorch_model.bin")
-        text_enc_dict = torch.load(text_enc_path, map_location="cpu")
-
-    # Convert the UNet model
-    unet_state_dict = diffusers_convert.convert_unet_state_dict(unet_state_dict)
-    unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()}
-
-    # Convert the VAE model
-    vae_state_dict = diffusers_convert.convert_vae_state_dict(vae_state_dict)
-    vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
-
-    # Easiest way to identify v2.0 model seems to be that the text encoder (OpenCLIP) is deeper
-    is_v20_model = "text_model.encoder.layers.22.layer_norm2.bias" in text_enc_dict
-
-    if is_v20_model:
-        # Need to add the tag 'transformer' in advance so we can knock it out from the final layer-norm
-        text_enc_dict = {"transformer." + k: v for k, v in text_enc_dict.items()}
-        text_enc_dict = diffusers_convert.convert_text_enc_state_dict_v20(text_enc_dict)
-        text_enc_dict = {"cond_stage_model.model." + k: v for k, v in text_enc_dict.items()}
-    else:
-        text_enc_dict = diffusers_convert.convert_text_enc_state_dict(text_enc_dict)
-        text_enc_dict = {"cond_stage_model.transformer." + k: v for k, v in text_enc_dict.items()}
-
-    # Put together new checkpoint
-    sd = {**unet_state_dict, **vae_state_dict, **text_enc_dict}
-
-    return load_checkpoint(embedding_directory=embedding_directory, state_dict=sd, config=config)
+    return (unet, clip, vae)
diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py
index 973619bf..34484b28 100644
--- a/comfy/ldm/modules/attention.py
+++ b/comfy/ldm/modules/attention.py
@@ -323,8 +323,7 @@ class CrossAttentionDoggettx(nn.Module):
                 break
             except model_management.OOM_EXCEPTION as e:
                 if first_op_done == False:
-                    torch.cuda.empty_cache()
-                    torch.cuda.ipc_collect()
+                    model_management.soft_empty_cache(True)
                     if cleared_cache == False:
                         cleared_cache = True
                         print("out of memory error, emptying cache and trying again")
@@ -402,8 +401,6 @@ class MemoryEfficientCrossAttention(nn.Module):
     # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
     def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0, dtype=None, device=None, operations=comfy.ops):
         super().__init__()
-        print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using "
-              f"{heads} heads.")
         inner_dim = dim_head * heads
         context_dim = default(context_dim, query_dim)
 
diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py
index b596408d..5f38640c 100644
--- a/comfy/ldm/modules/diffusionmodules/model.py
+++ b/comfy/ldm/modules/diffusionmodules/model.py
@@ -56,7 +56,18 @@ class Upsample(nn.Module):
                                         padding=1)
 
     def forward(self, x):
-        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        try:
+            x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        except: #operation not implemented for bf16
+            b, c, h, w = x.shape
+            out = torch.empty((b, c, h*2, w*2), dtype=x.dtype, layout=x.layout, device=x.device)
+            split = 8
+            l = out.shape[1] // split
+            for i in range(0, out.shape[1], l):
+                out[:,i:i+l] = torch.nn.functional.interpolate(x[:,i:i+l].to(torch.float32), scale_factor=2.0, mode="nearest").to(x.dtype)
+            del x
+            x = out
+
         if self.with_conv:
             x = self.conv(x)
         return x
@@ -74,11 +85,10 @@ class Downsample(nn.Module):
                                         stride=2,
                                         padding=0)
 
-    def forward(self, x, already_padded=False):
+    def forward(self, x):
         if self.with_conv:
-            if not already_padded:
-                pad = (0,1,0,1)
-                x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
             x = self.conv(x)
         else:
             x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
@@ -176,6 +186,7 @@ def slice_attention(q, k, v):
                 del s2
             break
         except model_management.OOM_EXCEPTION as e:
+            model_management.soft_empty_cache(True)
             steps *= 2
             if steps > 128:
                 raise e
@@ -275,25 +286,17 @@ class MemoryEfficientAttnBlock(nn.Module):
 
         # compute attention
         B, C, H, W = q.shape
-        q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v))
-
         q, k, v = map(
-            lambda t: t.unsqueeze(3)
-            .reshape(B, t.shape[1], 1, C)
-            .permute(0, 2, 1, 3)
-            .reshape(B * 1, t.shape[1], C)
-            .contiguous(),
+            lambda t: t.view(B, C, -1).transpose(1, 2).contiguous(),
             (q, k, v),
         )
-        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
 
-        out = (
-            out.unsqueeze(0)
-            .reshape(B, 1, out.shape[1], C)
-            .permute(0, 2, 1, 3)
-            .reshape(B, out.shape[1], C)
-        )
-        out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C)
+        try:
+            out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
+            out = out.transpose(1, 2).reshape(B, C, H, W)
+        except NotImplementedError as e:
+            out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(B, C, H, W)
+
         out = self.proj_out(out)
         return x+out
 
@@ -603,9 +606,6 @@ class Encoder(nn.Module):
     def forward(self, x):
         # timestep embedding
         temb = None
-        pad = (0,1,0,1)
-        x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
-        already_padded = True
         # downsampling
         h = self.conv_in(x)
         for i_level in range(self.num_resolutions):
@@ -614,8 +614,7 @@ class Encoder(nn.Module):
                 if len(self.down[i_level].attn) > 0:
                     h = self.down[i_level].attn[i_block](h)
             if i_level != self.num_resolutions-1:
-                h = self.down[i_level].downsample(h, already_padded)
-                already_padded = False
+                h = self.down[i_level].downsample(h)
 
         # middle
         h = self.mid.block_1(h, temb)
diff --git a/comfy/lora.py b/comfy/lora.py
new file mode 100644
index 00000000..3009a1c9
--- /dev/null
+++ b/comfy/lora.py
@@ -0,0 +1,199 @@
+import comfy.utils
+
+LORA_CLIP_MAP = {
+    "mlp.fc1": "mlp_fc1",
+    "mlp.fc2": "mlp_fc2",
+    "self_attn.k_proj": "self_attn_k_proj",
+    "self_attn.q_proj": "self_attn_q_proj",
+    "self_attn.v_proj": "self_attn_v_proj",
+    "self_attn.out_proj": "self_attn_out_proj",
+}
+
+
+def load_lora(lora, to_load):
+    patch_dict = {}
+    loaded_keys = set()
+    for x in to_load:
+        alpha_name = "{}.alpha".format(x)
+        alpha = None
+        if alpha_name in lora.keys():
+            alpha = lora[alpha_name].item()
+            loaded_keys.add(alpha_name)
+
+        regular_lora = "{}.lora_up.weight".format(x)
+        diffusers_lora = "{}_lora.up.weight".format(x)
+        transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
+        A_name = None
+
+        if regular_lora in lora.keys():
+            A_name = regular_lora
+            B_name = "{}.lora_down.weight".format(x)
+            mid_name = "{}.lora_mid.weight".format(x)
+        elif diffusers_lora in lora.keys():
+            A_name = diffusers_lora
+            B_name = "{}_lora.down.weight".format(x)
+            mid_name = None
+        elif transformers_lora in lora.keys():
+            A_name = transformers_lora
+            B_name ="{}.lora_linear_layer.down.weight".format(x)
+            mid_name = None
+
+        if A_name is not None:
+            mid = None
+            if mid_name is not None and mid_name in lora.keys():
+                mid = lora[mid_name]
+                loaded_keys.add(mid_name)
+            patch_dict[to_load[x]] = (lora[A_name], lora[B_name], alpha, mid)
+            loaded_keys.add(A_name)
+            loaded_keys.add(B_name)
+
+
+        ######## loha
+        hada_w1_a_name = "{}.hada_w1_a".format(x)
+        hada_w1_b_name = "{}.hada_w1_b".format(x)
+        hada_w2_a_name = "{}.hada_w2_a".format(x)
+        hada_w2_b_name = "{}.hada_w2_b".format(x)
+        hada_t1_name = "{}.hada_t1".format(x)
+        hada_t2_name = "{}.hada_t2".format(x)
+        if hada_w1_a_name in lora.keys():
+            hada_t1 = None
+            hada_t2 = None
+            if hada_t1_name in lora.keys():
+                hada_t1 = lora[hada_t1_name]
+                hada_t2 = lora[hada_t2_name]
+                loaded_keys.add(hada_t1_name)
+                loaded_keys.add(hada_t2_name)
+
+            patch_dict[to_load[x]] = (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2)
+            loaded_keys.add(hada_w1_a_name)
+            loaded_keys.add(hada_w1_b_name)
+            loaded_keys.add(hada_w2_a_name)
+            loaded_keys.add(hada_w2_b_name)
+
+
+        ######## lokr
+        lokr_w1_name = "{}.lokr_w1".format(x)
+        lokr_w2_name = "{}.lokr_w2".format(x)
+        lokr_w1_a_name = "{}.lokr_w1_a".format(x)
+        lokr_w1_b_name = "{}.lokr_w1_b".format(x)
+        lokr_t2_name = "{}.lokr_t2".format(x)
+        lokr_w2_a_name = "{}.lokr_w2_a".format(x)
+        lokr_w2_b_name = "{}.lokr_w2_b".format(x)
+
+        lokr_w1 = None
+        if lokr_w1_name in lora.keys():
+            lokr_w1 = lora[lokr_w1_name]
+            loaded_keys.add(lokr_w1_name)
+
+        lokr_w2 = None
+        if lokr_w2_name in lora.keys():
+            lokr_w2 = lora[lokr_w2_name]
+            loaded_keys.add(lokr_w2_name)
+
+        lokr_w1_a = None
+        if lokr_w1_a_name in lora.keys():
+            lokr_w1_a = lora[lokr_w1_a_name]
+            loaded_keys.add(lokr_w1_a_name)
+
+        lokr_w1_b = None
+        if lokr_w1_b_name in lora.keys():
+            lokr_w1_b = lora[lokr_w1_b_name]
+            loaded_keys.add(lokr_w1_b_name)
+
+        lokr_w2_a = None
+        if lokr_w2_a_name in lora.keys():
+            lokr_w2_a = lora[lokr_w2_a_name]
+            loaded_keys.add(lokr_w2_a_name)
+
+        lokr_w2_b = None
+        if lokr_w2_b_name in lora.keys():
+            lokr_w2_b = lora[lokr_w2_b_name]
+            loaded_keys.add(lokr_w2_b_name)
+
+        lokr_t2 = None
+        if lokr_t2_name in lora.keys():
+            lokr_t2 = lora[lokr_t2_name]
+            loaded_keys.add(lokr_t2_name)
+
+        if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
+            patch_dict[to_load[x]] = (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2)
+
+
+        w_norm_name = "{}.w_norm".format(x)
+        b_norm_name = "{}.b_norm".format(x)
+        w_norm = lora.get(w_norm_name, None)
+        b_norm = lora.get(b_norm_name, None)
+
+        if w_norm is not None:
+            loaded_keys.add(w_norm_name)
+            patch_dict[to_load[x]] = (w_norm,)
+            if b_norm is not None:
+                loaded_keys.add(b_norm_name)
+                patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = (b_norm,)
+
+    for x in lora.keys():
+        if x not in loaded_keys:
+            print("lora key not loaded", x)
+    return patch_dict
+
+def model_lora_keys_clip(model, key_map={}):
+    sdk = model.state_dict().keys()
+
+    text_model_lora_key = "lora_te_text_model_encoder_layers_{}_{}"
+    clip_l_present = False
+    for b in range(32):
+        for c in LORA_CLIP_MAP:
+            k = "transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
+                key_map[lora_key] = k
+                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c])
+                key_map[lora_key] = k
+                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                key_map[lora_key] = k
+
+            k = "clip_l.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
+                key_map[lora_key] = k
+                clip_l_present = True
+                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                key_map[lora_key] = k
+
+            k = "clip_g.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                if clip_l_present:
+                    lora_key = "lora_te2_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
+                    key_map[lora_key] = k
+                    lora_key = "text_encoder_2.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                    key_map[lora_key] = k
+                else:
+                    lora_key = "lora_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #TODO: test if this is correct for SDXL-Refiner
+                    key_map[lora_key] = k
+                    lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                    key_map[lora_key] = k
+
+    return key_map
+
+def model_lora_keys_unet(model, key_map={}):
+    sdk = model.state_dict().keys()
+
+    for k in sdk:
+        if k.startswith("diffusion_model.") and k.endswith(".weight"):
+            key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
+            key_map["lora_unet_{}".format(key_lora)] = k
+
+    diffusers_keys = comfy.utils.unet_to_diffusers(model.model_config.unet_config)
+    for k in diffusers_keys:
+        if k.endswith(".weight"):
+            unet_key = "diffusion_model.{}".format(diffusers_keys[k])
+            key_lora = k[:-len(".weight")].replace(".", "_")
+            key_map["lora_unet_{}".format(key_lora)] = unet_key
+
+            diffusers_lora_prefix = ["", "unet."]
+            for p in diffusers_lora_prefix:
+                diffusers_lora_key = "{}{}".format(p, k[:-len(".weight")].replace(".to_", ".processor.to_"))
+                if diffusers_lora_key.endswith(".to_out.0"):
+                    diffusers_lora_key = diffusers_lora_key[:-2]
+                key_map[diffusers_lora_key] = unet_key
+    return key_map
diff --git a/comfy/model_base.py b/comfy/model_base.py
index 979e2c65..ca154dba 100644
--- a/comfy/model_base.py
+++ b/comfy/model_base.py
@@ -3,6 +3,7 @@ from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel
 from comfy.ldm.modules.encoders.noise_aug_modules import CLIPEmbeddingNoiseAugmentation
 from comfy.ldm.modules.diffusionmodules.util import make_beta_schedule
 from comfy.ldm.modules.diffusionmodules.openaimodel import Timestep
+import comfy.model_management
 import numpy as np
 from enum import Enum
 from . import utils
@@ -18,8 +19,9 @@ class BaseModel(torch.nn.Module):
         unet_config = model_config.unet_config
         self.latent_format = model_config.latent_format
         self.model_config = model_config
-        self.register_schedule(given_betas=None, beta_schedule="linear", timesteps=1000, linear_start=0.00085, linear_end=0.012, cosine_s=8e-3)
-        self.diffusion_model = UNetModel(**unet_config, device=device)
+        self.register_schedule(given_betas=None, beta_schedule=model_config.beta_schedule, timesteps=1000, linear_start=0.00085, linear_end=0.012, cosine_s=8e-3)
+        if not unet_config.get("disable_unet_model_creation", False):
+            self.diffusion_model = UNetModel(**unet_config, device=device)
         self.model_type = model_type
         self.adm_channels = unet_config.get("adm_in_channels", None)
         if self.adm_channels is None:
@@ -48,10 +50,10 @@ class BaseModel(torch.nn.Module):
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, c_adm=None, control=None, transformer_options={}):
         if c_concat is not None:
-            xc = torch.cat([x] + c_concat, dim=1)
+            xc = torch.cat([x] + [c_concat], dim=1)
         else:
             xc = x
-        context = torch.cat(c_crossattn, 1)
+        context = c_crossattn
         dtype = self.get_dtype()
         xc = xc.to(dtype)
         t = t.to(dtype)
@@ -93,7 +95,11 @@ class BaseModel(torch.nn.Module):
 
     def state_dict_for_saving(self, clip_state_dict, vae_state_dict):
         clip_state_dict = self.model_config.process_clip_state_dict_for_saving(clip_state_dict)
-        unet_state_dict = self.diffusion_model.state_dict()
+        unet_sd = self.diffusion_model.state_dict()
+        unet_state_dict = {}
+        for k in unet_sd:
+            unet_state_dict[k] = comfy.model_management.resolve_lowvram_weight(unet_sd[k], self.diffusion_model, k)
+
         unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
         vae_state_dict = self.model_config.process_vae_state_dict_for_saving(vae_state_dict)
         if self.get_dtype() == torch.float16:
@@ -105,6 +111,9 @@ class BaseModel(torch.nn.Module):
 
         return {**unet_state_dict, **vae_state_dict, **clip_state_dict}
 
+    def set_inpaint(self):
+        self.concat_keys = ("mask", "masked_image")
+
 def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0):
     adm_inputs = []
     weights = []
@@ -142,12 +151,6 @@ class SD21UNCLIP(BaseModel):
         else:
             return unclip_adm(unclip_conditioning, device, self.noise_augmentor, kwargs.get("unclip_noise_augment_merge", 0.05))
 
-
-class SDInpaint(BaseModel):
-    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
-        super().__init__(model_config, model_type, device=device)
-        self.concat_keys = ("mask", "masked_image")
-
 def sdxl_pooled(args, noise_augmentor):
     if "unclip_conditioning" in args:
         return unclip_adm(args.get("unclip_conditioning", None), args["device"], noise_augmentor)[:,:1280]
diff --git a/comfy/model_detection.py b/comfy/model_detection.py
index 0edc4f18..372d5a2d 100644
--- a/comfy/model_detection.py
+++ b/comfy/model_detection.py
@@ -183,8 +183,12 @@ def unet_config_from_diffusers_unet(state_dict, use_fp16):
             'num_res_blocks': 2, 'attention_resolutions': [], 'transformer_depth': [0, 0, 0], 'channel_mult': [1, 2, 4],
             'transformer_depth_middle': 0, 'use_linear_in_transformer': True, "num_head_channels": 64, 'context_dim': 1}
 
+    SDXL_diffusers_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'num_classes': 'sequential', 'adm_in_channels': 2816, 'use_fp16': use_fp16, 'in_channels': 9, 'model_channels': 320,
+            'num_res_blocks': 2, 'attention_resolutions': [2, 4], 'transformer_depth': [0, 2, 10], 'channel_mult': [1, 2, 4],
+            'transformer_depth_middle': 10, 'use_linear_in_transformer': True, 'context_dim': 2048, "num_head_channels": 64}
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet]
+    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint]
 
     for unet_config in supported_models:
         matches = True
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 0e86df41..b663e8f5 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -1,6 +1,7 @@
 import psutil
 from enum import Enum
 from comfy.cli_args import args
+import comfy.utils
 import torch
 import sys
 
@@ -57,8 +58,15 @@ except:
 if args.cpu:
     cpu_state = CPUState.CPU
 
-def get_torch_device():
+def is_intel_xpu():
+    global cpu_state
     global xpu_available
+    if cpu_state == CPUState.GPU:
+        if xpu_available:
+            return True
+    return False
+
+def get_torch_device():
     global directml_enabled
     global cpu_state
     if directml_enabled:
@@ -69,13 +77,12 @@ def get_torch_device():
     if cpu_state == CPUState.CPU:
         return torch.device("cpu")
     else:
-        if xpu_available:
+        if is_intel_xpu():
             return torch.device("xpu")
         else:
             return torch.device(torch.cuda.current_device())
 
 def get_total_memory(dev=None, torch_total_too=False):
-    global xpu_available
     global directml_enabled
     if dev is None:
         dev = get_torch_device()
@@ -87,7 +94,7 @@ def get_total_memory(dev=None, torch_total_too=False):
         if directml_enabled:
             mem_total = 1024 * 1024 * 1024 #TODO
             mem_total_torch = mem_total
-        elif xpu_available:
+        elif is_intel_xpu():
             stats = torch.xpu.memory_stats(dev)
             mem_reserved = stats['reserved_bytes.all.current']
             mem_total = torch.xpu.get_device_properties(dev).total_memory
@@ -111,9 +118,6 @@ if not args.normalvram and not args.cpu:
     if lowvram_available and total_vram <= 4096:
         print("Trying to enable lowvram mode because your GPU seems to have 4GB or less. If you don't want this use: --normalvram")
         set_vram_to = VRAMState.LOW_VRAM
-    elif total_vram > total_ram * 1.1 and total_vram > 14336:
-        print("Enabling highvram mode because your GPU has more vram than your computer has ram. If you don't want this use: --normalvram")
-        vram_state = VRAMState.HIGH_VRAM
 
 try:
     OOM_EXCEPTION = torch.cuda.OutOfMemoryError
@@ -148,17 +152,32 @@ def is_nvidia():
     if cpu_state == CPUState.GPU:
         if torch.version.cuda:
             return True
+    return False
 
 ENABLE_PYTORCH_ATTENTION = args.use_pytorch_cross_attention
+VAE_DTYPE = torch.float32
 
-if ENABLE_PYTORCH_ATTENTION == False and XFORMERS_IS_AVAILABLE == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
-    try:
-        if is_nvidia():
-            torch_version = torch.version.__version__
-            if int(torch_version[0]) >= 2:
+try:
+    if is_nvidia():
+        torch_version = torch.version.__version__
+        if int(torch_version[0]) >= 2:
+            if ENABLE_PYTORCH_ATTENTION == False and XFORMERS_IS_AVAILABLE == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
-    except:
-        pass
+            if torch.cuda.is_bf16_supported():
+                VAE_DTYPE = torch.bfloat16
+except:
+    pass
+
+if is_intel_xpu():
+    VAE_DTYPE = torch.bfloat16
+
+if args.fp16_vae:
+    VAE_DTYPE = torch.float16
+elif args.bf16_vae:
+    VAE_DTYPE = torch.bfloat16
+elif args.fp32_vae:
+    VAE_DTYPE = torch.float32
+
 
 if ENABLE_PYTORCH_ATTENTION:
     torch.backends.cuda.enable_math_sdp(True)
@@ -210,7 +229,6 @@ if DISABLE_SMART_MEMORY:
     print("Disabling smart memory management")
 
 def get_torch_device_name(device):
-    global xpu_available
     if hasattr(device, 'type'):
         if device.type == "cuda":
             try:
@@ -220,7 +238,7 @@ def get_torch_device_name(device):
             return "{} {} : {}".format(device, torch.cuda.get_device_name(device), allocator_backend)
         else:
             return "{}".format(device.type)
-    elif xpu_available:
+    elif is_intel_xpu():
         return "{} {}".format(device, torch.xpu.get_device_name(device))
     else:
         return "CUDA {}: {}".format(device, torch.cuda.get_device_name(device))
@@ -230,6 +248,7 @@ try:
 except:
     print("Could not pick default device.")
 
+print("VAE dtype:", VAE_DTYPE)
 
 current_loaded_models = []
 
@@ -249,7 +268,6 @@ class LoadedModel:
             return self.model_memory()
 
     def model_load(self, lowvram_model_memory=0):
-        global xpu_available
         patch_model_to = None
         if lowvram_model_memory == 0:
             patch_model_to = self.device
@@ -270,7 +288,7 @@ class LoadedModel:
             accelerate.dispatch_model(self.real_model, device_map=device_map, main_device=self.device)
             self.model_accelerated = True
 
-        if xpu_available and not args.disable_ipex_optimize:
+        if is_intel_xpu() and not args.disable_ipex_optimize:
             self.real_model = torch.xpu.optimize(self.real_model.eval(), inplace=True, auto_kernel_selection=True, graph_mode=True)
 
         return self.real_model
@@ -450,12 +468,8 @@ def vae_offload_device():
         return torch.device("cpu")
 
 def vae_dtype():
-    if args.fp16_vae:
-        return torch.float16
-    elif args.bf16_vae:
-        return torch.bfloat16
-    else:
-        return torch.float32
+    global VAE_DTYPE
+    return VAE_DTYPE
 
 def get_autocast_device(dev):
     if hasattr(dev, 'type'):
@@ -464,12 +478,11 @@ def get_autocast_device(dev):
 
 
 def xformers_enabled():
-    global xpu_available
     global directml_enabled
     global cpu_state
     if cpu_state != CPUState.GPU:
         return False
-    if xpu_available:
+    if is_intel_xpu():
         return False
     if directml_enabled:
         return False
@@ -496,7 +509,6 @@ def pytorch_attention_flash_attention():
     return False
 
 def get_free_memory(dev=None, torch_free_too=False):
-    global xpu_available
     global directml_enabled
     if dev is None:
         dev = get_torch_device()
@@ -508,7 +520,7 @@ def get_free_memory(dev=None, torch_free_too=False):
         if directml_enabled:
             mem_free_total = 1024 * 1024 * 1024 #TODO
             mem_free_torch = mem_free_total
-        elif xpu_available:
+        elif is_intel_xpu():
             stats = torch.xpu.memory_stats(dev)
             mem_active = stats['active_bytes.all.current']
             mem_allocated = stats['allocated_bytes.all.current']
@@ -570,7 +582,6 @@ def is_device_mps(device):
     return False
 
 def should_use_fp16(device=None, model_params=0, prioritize_performance=True):
-    global xpu_available
     global directml_enabled
 
     if device is not None:
@@ -593,7 +604,7 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True):
     if cpu_mode() or mps_mode():
         return False #TODO ?
 
-    if xpu_available:
+    if is_intel_xpu():
         return True
 
     if torch.cuda.is_bf16_supported():
@@ -628,18 +639,24 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True):
 
     return True
 
-def soft_empty_cache():
-    global xpu_available
+def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:
         torch.mps.empty_cache()
-    elif xpu_available:
+    elif is_intel_xpu():
         torch.xpu.empty_cache()
     elif torch.cuda.is_available():
-        if is_nvidia(): #This seems to make things worse on ROCm so I only do it for cuda
+        if force or is_nvidia(): #This seems to make things worse on ROCm so I only do it for cuda
             torch.cuda.empty_cache()
             torch.cuda.ipc_collect()
 
+def resolve_lowvram_weight(weight, model, key):
+    if weight.device == torch.device("meta"): #lowvram NOTE: this depends on the inner working of the accelerate library so it might break.
+        key_split = key.split('.')              # I have no idea why they don't just leave the weight there instead of using the meta device.
+        op = comfy.utils.get_attr(model, '.'.join(key_split[:-1]))
+        weight = op._hf_hook.weights_map[key_split[-1]]
+    return weight
+
 #TODO: might be cleaner to put this somewhere else
 import threading
 
diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py
new file mode 100644
index 00000000..a6ee0bae
--- /dev/null
+++ b/comfy/model_patcher.py
@@ -0,0 +1,270 @@
+import torch
+import copy
+import inspect
+
+import comfy.utils
+
+class ModelPatcher:
+    def __init__(self, model, load_device, offload_device, size=0, current_device=None):
+        self.size = size
+        self.model = model
+        self.patches = {}
+        self.backup = {}
+        self.model_options = {"transformer_options":{}}
+        self.model_size()
+        self.load_device = load_device
+        self.offload_device = offload_device
+        if current_device is None:
+            self.current_device = self.offload_device
+        else:
+            self.current_device = current_device
+
+    def model_size(self):
+        if self.size > 0:
+            return self.size
+        model_sd = self.model.state_dict()
+        size = 0
+        for k in model_sd:
+            t = model_sd[k]
+            size += t.nelement() * t.element_size()
+        self.size = size
+        self.model_keys = set(model_sd.keys())
+        return size
+
+    def clone(self):
+        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, self.current_device)
+        n.patches = {}
+        for k in self.patches:
+            n.patches[k] = self.patches[k][:]
+
+        n.model_options = copy.deepcopy(self.model_options)
+        n.model_keys = self.model_keys
+        return n
+
+    def is_clone(self, other):
+        if hasattr(other, 'model') and self.model is other.model:
+            return True
+        return False
+
+    def set_model_sampler_cfg_function(self, sampler_cfg_function):
+        if len(inspect.signature(sampler_cfg_function).parameters) == 3:
+            self.model_options["sampler_cfg_function"] = lambda args: sampler_cfg_function(args["cond"], args["uncond"], args["cond_scale"]) #Old way
+        else:
+            self.model_options["sampler_cfg_function"] = sampler_cfg_function
+
+    def set_model_unet_function_wrapper(self, unet_wrapper_function):
+        self.model_options["model_function_wrapper"] = unet_wrapper_function
+
+    def set_model_patch(self, patch, name):
+        to = self.model_options["transformer_options"]
+        if "patches" not in to:
+            to["patches"] = {}
+        to["patches"][name] = to["patches"].get(name, []) + [patch]
+
+    def set_model_patch_replace(self, patch, name, block_name, number):
+        to = self.model_options["transformer_options"]
+        if "patches_replace" not in to:
+            to["patches_replace"] = {}
+        if name not in to["patches_replace"]:
+            to["patches_replace"][name] = {}
+        to["patches_replace"][name][(block_name, number)] = patch
+
+    def set_model_attn1_patch(self, patch):
+        self.set_model_patch(patch, "attn1_patch")
+
+    def set_model_attn2_patch(self, patch):
+        self.set_model_patch(patch, "attn2_patch")
+
+    def set_model_attn1_replace(self, patch, block_name, number):
+        self.set_model_patch_replace(patch, "attn1", block_name, number)
+
+    def set_model_attn2_replace(self, patch, block_name, number):
+        self.set_model_patch_replace(patch, "attn2", block_name, number)
+
+    def set_model_attn1_output_patch(self, patch):
+        self.set_model_patch(patch, "attn1_output_patch")
+
+    def set_model_attn2_output_patch(self, patch):
+        self.set_model_patch(patch, "attn2_output_patch")
+
+    def model_patches_to(self, device):
+        to = self.model_options["transformer_options"]
+        if "patches" in to:
+            patches = to["patches"]
+            for name in patches:
+                patch_list = patches[name]
+                for i in range(len(patch_list)):
+                    if hasattr(patch_list[i], "to"):
+                        patch_list[i] = patch_list[i].to(device)
+        if "patches_replace" in to:
+            patches = to["patches_replace"]
+            for name in patches:
+                patch_list = patches[name]
+                for k in patch_list:
+                    if hasattr(patch_list[k], "to"):
+                        patch_list[k] = patch_list[k].to(device)
+
+    def model_dtype(self):
+        if hasattr(self.model, "get_dtype"):
+            return self.model.get_dtype()
+
+    def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
+        p = set()
+        for k in patches:
+            if k in self.model_keys:
+                p.add(k)
+                current_patches = self.patches.get(k, [])
+                current_patches.append((strength_patch, patches[k], strength_model))
+                self.patches[k] = current_patches
+
+        return list(p)
+
+    def get_key_patches(self, filter_prefix=None):
+        model_sd = self.model_state_dict()
+        p = {}
+        for k in model_sd:
+            if filter_prefix is not None:
+                if not k.startswith(filter_prefix):
+                    continue
+            if k in self.patches:
+                p[k] = [model_sd[k]] + self.patches[k]
+            else:
+                p[k] = (model_sd[k],)
+        return p
+
+    def model_state_dict(self, filter_prefix=None):
+        sd = self.model.state_dict()
+        keys = list(sd.keys())
+        if filter_prefix is not None:
+            for k in keys:
+                if not k.startswith(filter_prefix):
+                    sd.pop(k)
+        return sd
+
+    def patch_model(self, device_to=None):
+        model_sd = self.model_state_dict()
+        for key in self.patches:
+            if key not in model_sd:
+                print("could not patch. key doesn't exist in model:", key)
+                continue
+
+            weight = model_sd[key]
+
+            if key not in self.backup:
+                self.backup[key] = weight.to(self.offload_device)
+
+            if device_to is not None:
+                temp_weight = weight.float().to(device_to, copy=True)
+            else:
+                temp_weight = weight.to(torch.float32, copy=True)
+            out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype)
+            comfy.utils.set_attr(self.model, key, out_weight)
+            del temp_weight
+
+        if device_to is not None:
+            self.model.to(device_to)
+            self.current_device = device_to
+
+        return self.model
+
+    def calculate_weight(self, patches, weight, key):
+        for p in patches:
+            alpha = p[0]
+            v = p[1]
+            strength_model = p[2]
+
+            if strength_model != 1.0:
+                weight *= strength_model
+
+            if isinstance(v, list):
+                v = (self.calculate_weight(v[1:], v[0].clone(), key), )
+
+            if len(v) == 1:
+                w1 = v[0]
+                if alpha != 0.0:
+                    if w1.shape != weight.shape:
+                        print("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, w1.shape, weight.shape))
+                    else:
+                        weight += alpha * w1.type(weight.dtype).to(weight.device)
+            elif len(v) == 4: #lora/locon
+                mat1 = v[0].float().to(weight.device)
+                mat2 = v[1].float().to(weight.device)
+                if v[2] is not None:
+                    alpha *= v[2] / mat2.shape[0]
+                if v[3] is not None:
+                    #locon mid weights, hopefully the math is fine because I didn't properly test it
+                    mat3 = v[3].float().to(weight.device)
+                    final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
+                    mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
+                try:
+                    weight += (alpha * torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1))).reshape(weight.shape).type(weight.dtype)
+                except Exception as e:
+                    print("ERROR", key, e)
+            elif len(v) == 8: #lokr
+                w1 = v[0]
+                w2 = v[1]
+                w1_a = v[3]
+                w1_b = v[4]
+                w2_a = v[5]
+                w2_b = v[6]
+                t2 = v[7]
+                dim = None
+
+                if w1 is None:
+                    dim = w1_b.shape[0]
+                    w1 = torch.mm(w1_a.float(), w1_b.float())
+                else:
+                    w1 = w1.float().to(weight.device)
+
+                if w2 is None:
+                    dim = w2_b.shape[0]
+                    if t2 is None:
+                        w2 = torch.mm(w2_a.float().to(weight.device), w2_b.float().to(weight.device))
+                    else:
+                        w2 = torch.einsum('i j k l, j r, i p -> p r k l', t2.float().to(weight.device), w2_b.float().to(weight.device), w2_a.float().to(weight.device))
+                else:
+                    w2 = w2.float().to(weight.device)
+
+                if len(w2.shape) == 4:
+                    w1 = w1.unsqueeze(2).unsqueeze(2)
+                if v[2] is not None and dim is not None:
+                    alpha *= v[2] / dim
+
+                try:
+                    weight += alpha * torch.kron(w1, w2).reshape(weight.shape).type(weight.dtype)
+                except Exception as e:
+                    print("ERROR", key, e)
+            else: #loha
+                w1a = v[0]
+                w1b = v[1]
+                if v[2] is not None:
+                    alpha *= v[2] / w1b.shape[0]
+                w2a = v[3]
+                w2b = v[4]
+                if v[5] is not None: #cp decomposition
+                    t1 = v[5]
+                    t2 = v[6]
+                    m1 = torch.einsum('i j k l, j r, i p -> p r k l', t1.float().to(weight.device), w1b.float().to(weight.device), w1a.float().to(weight.device))
+                    m2 = torch.einsum('i j k l, j r, i p -> p r k l', t2.float().to(weight.device), w2b.float().to(weight.device), w2a.float().to(weight.device))
+                else:
+                    m1 = torch.mm(w1a.float().to(weight.device), w1b.float().to(weight.device))
+                    m2 = torch.mm(w2a.float().to(weight.device), w2b.float().to(weight.device))
+
+                try:
+                    weight += (alpha * m1 * m2).reshape(weight.shape).type(weight.dtype)
+                except Exception as e:
+                    print("ERROR", key, e)
+
+        return weight
+
+    def unpatch_model(self, device_to=None):
+        keys = list(self.backup.keys())
+
+        for k in keys:
+            comfy.utils.set_attr(self.model, k, self.backup[k])
+
+        self.backup = {}
+
+        if device_to is not None:
+            self.model.to(device_to)
+            self.current_device = device_to
diff --git a/comfy/sample.py b/comfy/sample.py
index 79ea37e0..e4730b18 100644
--- a/comfy/sample.py
+++ b/comfy/sample.py
@@ -1,6 +1,7 @@
 import torch
 import comfy.model_management
 import comfy.samplers
+import comfy.utils
 import math
 import numpy as np
 
@@ -28,8 +29,7 @@ def prepare_mask(noise_mask, shape, device):
     noise_mask = torch.nn.functional.interpolate(noise_mask.reshape((-1, 1, noise_mask.shape[-2], noise_mask.shape[-1])), size=(shape[2], shape[3]), mode="bilinear")
     noise_mask = noise_mask.round()
     noise_mask = torch.cat([noise_mask] * shape[1], dim=1)
-    if noise_mask.shape[0] < shape[0]:
-        noise_mask = noise_mask.repeat(math.ceil(shape[0] / noise_mask.shape[0]), 1, 1, 1)[:shape[0]]
+    noise_mask = comfy.utils.repeat_to_batch_size(noise_mask, shape[0])
     noise_mask = noise_mask.to(device)
     return noise_mask
 
@@ -37,9 +37,7 @@ def broadcast_cond(cond, batch, device):
     """broadcasts conditioning to the batch size"""
     copy = []
     for p in cond:
-        t = p[0]
-        if t.shape[0] < batch:
-            t = torch.cat([t] * batch)
+        t = comfy.utils.repeat_to_batch_size(p[0], batch)
         t = t.to(device)
         copy += [[t] + p[1:]]
     return copy
diff --git a/comfy/samplers.py b/comfy/samplers.py
index 134336de..3250b2ed 100644
--- a/comfy/samplers.py
+++ b/comfy/samplers.py
@@ -165,9 +165,9 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, con
                 c_crossattn_out.append(c)
 
             if len(c_crossattn_out) > 0:
-                out['c_crossattn'] = [torch.cat(c_crossattn_out)]
+                out['c_crossattn'] = torch.cat(c_crossattn_out)
             if len(c_concat) > 0:
-                out['c_concat'] = [torch.cat(c_concat)]
+                out['c_concat'] = torch.cat(c_concat)
             if len(c_adm) > 0:
                 out['c_adm'] = torch.cat(c_adm)
             return out
@@ -390,11 +390,20 @@ def get_mask_aabb(masks):
 
     return bounding_boxes, is_empty
 
-def resolve_cond_masks(conditions, h, w, device):
+def resolve_areas_and_cond_masks(conditions, h, w, device):
     # We need to decide on an area outside the sampling loop in order to properly generate opposite areas of equal sizes.
     # While we're doing this, we can also resolve the mask device and scaling for performance reasons
     for i in range(len(conditions)):
         c = conditions[i]
+        if 'area' in c[1]:
+            area = c[1]['area']
+            if area[0] == "percentage":
+                modified = c[1].copy()
+                area = (max(1, round(area[1] * h)), max(1, round(area[2] * w)), round(area[3] * h), round(area[4] * w))
+                modified['area'] = area
+                c = [c[0], modified]
+                conditions[i] = c
+
         if 'mask' in c[1]:
             mask = c[1]['mask']
             mask = mask.to(device=device)
@@ -622,8 +631,8 @@ class KSampler:
         positive = positive[:]
         negative = negative[:]
 
-        resolve_cond_masks(positive, noise.shape[2], noise.shape[3], self.device)
-        resolve_cond_masks(negative, noise.shape[2], noise.shape[3], self.device)
+        resolve_areas_and_cond_masks(positive, noise.shape[2], noise.shape[3], self.device)
+        resolve_areas_and_cond_masks(negative, noise.shape[2], noise.shape[3], self.device)
 
         calculate_start_end_timesteps(self.model_wrap, negative)
         calculate_start_end_timesteps(self.model_wrap, positive)
diff --git a/comfy/sd.py b/comfy/sd.py
index 3568a2aa..8be0bcbc 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -1,17 +1,14 @@
 import torch
 import contextlib
-import copy
-import inspect
 import math
 
 from comfy import model_management
 from .ldm.util import instantiate_from_config
 from .ldm.models.autoencoder import AutoencoderKL
 import yaml
-from .cldm import cldm
-from .t2i_adapter import adapter
 
-from . import utils
+import comfy.utils
+
 from . import clip_vision
 from . import gligen
 from . import diffusers_convert
@@ -22,6 +19,11 @@ from . import sd1_clip
 from . import sd2_clip
 from . import sdxl_clip
 
+import comfy.model_patcher
+import comfy.lora
+import comfy.t2i_adapter.adapter
+import comfy.supported_models_base
+
 def load_model_weights(model, sd):
     m, u = model.load_state_dict(sd, strict=False)
     m = set(m)
@@ -48,479 +50,14 @@ def load_clip_weights(model, sd):
         if ids.dtype == torch.float32:
             sd['cond_stage_model.transformer.text_model.embeddings.position_ids'] = ids.round()
 
-    sd = utils.transformers_convert(sd, "cond_stage_model.model.", "cond_stage_model.transformer.text_model.", 24)
+    sd = comfy.utils.transformers_convert(sd, "cond_stage_model.model.", "cond_stage_model.transformer.text_model.", 24)
     return load_model_weights(model, sd)
 
-LORA_CLIP_MAP = {
-    "mlp.fc1": "mlp_fc1",
-    "mlp.fc2": "mlp_fc2",
-    "self_attn.k_proj": "self_attn_k_proj",
-    "self_attn.q_proj": "self_attn_q_proj",
-    "self_attn.v_proj": "self_attn_v_proj",
-    "self_attn.out_proj": "self_attn_out_proj",
-}
-
-
-def load_lora(lora, to_load):
-    patch_dict = {}
-    loaded_keys = set()
-    for x in to_load:
-        alpha_name = "{}.alpha".format(x)
-        alpha = None
-        if alpha_name in lora.keys():
-            alpha = lora[alpha_name].item()
-            loaded_keys.add(alpha_name)
-
-        regular_lora = "{}.lora_up.weight".format(x)
-        diffusers_lora = "{}_lora.up.weight".format(x)
-        transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
-        A_name = None
-
-        if regular_lora in lora.keys():
-            A_name = regular_lora
-            B_name = "{}.lora_down.weight".format(x)
-            mid_name = "{}.lora_mid.weight".format(x)
-        elif diffusers_lora in lora.keys():
-            A_name = diffusers_lora
-            B_name = "{}_lora.down.weight".format(x)
-            mid_name = None
-        elif transformers_lora in lora.keys():
-            A_name = transformers_lora
-            B_name ="{}.lora_linear_layer.down.weight".format(x)
-            mid_name = None
-
-        if A_name is not None:
-            mid = None
-            if mid_name is not None and mid_name in lora.keys():
-                mid = lora[mid_name]
-                loaded_keys.add(mid_name)
-            patch_dict[to_load[x]] = (lora[A_name], lora[B_name], alpha, mid)
-            loaded_keys.add(A_name)
-            loaded_keys.add(B_name)
-
-
-        ######## loha
-        hada_w1_a_name = "{}.hada_w1_a".format(x)
-        hada_w1_b_name = "{}.hada_w1_b".format(x)
-        hada_w2_a_name = "{}.hada_w2_a".format(x)
-        hada_w2_b_name = "{}.hada_w2_b".format(x)
-        hada_t1_name = "{}.hada_t1".format(x)
-        hada_t2_name = "{}.hada_t2".format(x)
-        if hada_w1_a_name in lora.keys():
-            hada_t1 = None
-            hada_t2 = None
-            if hada_t1_name in lora.keys():
-                hada_t1 = lora[hada_t1_name]
-                hada_t2 = lora[hada_t2_name]
-                loaded_keys.add(hada_t1_name)
-                loaded_keys.add(hada_t2_name)
-
-            patch_dict[to_load[x]] = (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2)
-            loaded_keys.add(hada_w1_a_name)
-            loaded_keys.add(hada_w1_b_name)
-            loaded_keys.add(hada_w2_a_name)
-            loaded_keys.add(hada_w2_b_name)
-
-
-        ######## lokr
-        lokr_w1_name = "{}.lokr_w1".format(x)
-        lokr_w2_name = "{}.lokr_w2".format(x)
-        lokr_w1_a_name = "{}.lokr_w1_a".format(x)
-        lokr_w1_b_name = "{}.lokr_w1_b".format(x)
-        lokr_t2_name = "{}.lokr_t2".format(x)
-        lokr_w2_a_name = "{}.lokr_w2_a".format(x)
-        lokr_w2_b_name = "{}.lokr_w2_b".format(x)
-
-        lokr_w1 = None
-        if lokr_w1_name in lora.keys():
-            lokr_w1 = lora[lokr_w1_name]
-            loaded_keys.add(lokr_w1_name)
-
-        lokr_w2 = None
-        if lokr_w2_name in lora.keys():
-            lokr_w2 = lora[lokr_w2_name]
-            loaded_keys.add(lokr_w2_name)
-
-        lokr_w1_a = None
-        if lokr_w1_a_name in lora.keys():
-            lokr_w1_a = lora[lokr_w1_a_name]
-            loaded_keys.add(lokr_w1_a_name)
-
-        lokr_w1_b = None
-        if lokr_w1_b_name in lora.keys():
-            lokr_w1_b = lora[lokr_w1_b_name]
-            loaded_keys.add(lokr_w1_b_name)
-
-        lokr_w2_a = None
-        if lokr_w2_a_name in lora.keys():
-            lokr_w2_a = lora[lokr_w2_a_name]
-            loaded_keys.add(lokr_w2_a_name)
-
-        lokr_w2_b = None
-        if lokr_w2_b_name in lora.keys():
-            lokr_w2_b = lora[lokr_w2_b_name]
-            loaded_keys.add(lokr_w2_b_name)
-
-        lokr_t2 = None
-        if lokr_t2_name in lora.keys():
-            lokr_t2 = lora[lokr_t2_name]
-            loaded_keys.add(lokr_t2_name)
-
-        if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
-            patch_dict[to_load[x]] = (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2)
-
-    for x in lora.keys():
-        if x not in loaded_keys:
-            print("lora key not loaded", x)
-    return patch_dict
-
-def model_lora_keys_clip(model, key_map={}):
-    sdk = model.state_dict().keys()
-
-    text_model_lora_key = "lora_te_text_model_encoder_layers_{}_{}"
-    clip_l_present = False
-    for b in range(32):
-        for c in LORA_CLIP_MAP:
-            k = "transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
-            if k in sdk:
-                lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
-                key_map[lora_key] = k
-                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c])
-                key_map[lora_key] = k
-                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
-                key_map[lora_key] = k
-
-            k = "clip_l.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
-            if k in sdk:
-                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
-                key_map[lora_key] = k
-                clip_l_present = True
-                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
-                key_map[lora_key] = k
-
-            k = "clip_g.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
-            if k in sdk:
-                if clip_l_present:
-                    lora_key = "lora_te2_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
-                    key_map[lora_key] = k
-                    lora_key = "text_encoder_2.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
-                    key_map[lora_key] = k
-                else:
-                    lora_key = "lora_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #TODO: test if this is correct for SDXL-Refiner
-                    key_map[lora_key] = k
-                    lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
-                    key_map[lora_key] = k
-
-    return key_map
-
-def model_lora_keys_unet(model, key_map={}):
-    sdk = model.state_dict().keys()
-
-    for k in sdk:
-        if k.startswith("diffusion_model.") and k.endswith(".weight"):
-            key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
-            key_map["lora_unet_{}".format(key_lora)] = k
-
-    diffusers_keys = utils.unet_to_diffusers(model.model_config.unet_config)
-    for k in diffusers_keys:
-        if k.endswith(".weight"):
-            unet_key = "diffusion_model.{}".format(diffusers_keys[k])
-            key_lora = k[:-len(".weight")].replace(".", "_")
-            key_map["lora_unet_{}".format(key_lora)] = unet_key
-
-            diffusers_lora_prefix = ["", "unet."]
-            for p in diffusers_lora_prefix:
-                diffusers_lora_key = "{}{}".format(p, k[:-len(".weight")].replace(".to_", ".processor.to_"))
-                if diffusers_lora_key.endswith(".to_out.0"):
-                    diffusers_lora_key = diffusers_lora_key[:-2]
-                key_map[diffusers_lora_key] = unet_key
-    return key_map
-
-def set_attr(obj, attr, value):
-    attrs = attr.split(".")
-    for name in attrs[:-1]:
-        obj = getattr(obj, name)
-    prev = getattr(obj, attrs[-1])
-    setattr(obj, attrs[-1], torch.nn.Parameter(value))
-    del prev
-
-def get_attr(obj, attr):
-    attrs = attr.split(".")
-    for name in attrs:
-        obj = getattr(obj, name)
-    return obj
-
-
-class ModelPatcher:
-    def __init__(self, model, load_device, offload_device, size=0, current_device=None):
-        self.size = size
-        self.model = model
-        self.patches = {}
-        self.backup = {}
-        self.model_options = {"transformer_options":{}}
-        self.model_size()
-        self.load_device = load_device
-        self.offload_device = offload_device
-        if current_device is None:
-            self.current_device = self.offload_device
-        else:
-            self.current_device = current_device
-
-    def model_size(self):
-        if self.size > 0:
-            return self.size
-        model_sd = self.model.state_dict()
-        size = 0
-        for k in model_sd:
-            t = model_sd[k]
-            size += t.nelement() * t.element_size()
-        self.size = size
-        self.model_keys = set(model_sd.keys())
-        return size
-
-    def clone(self):
-        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, self.current_device)
-        n.patches = {}
-        for k in self.patches:
-            n.patches[k] = self.patches[k][:]
-
-        n.model_options = copy.deepcopy(self.model_options)
-        n.model_keys = self.model_keys
-        return n
-
-    def is_clone(self, other):
-        if hasattr(other, 'model') and self.model is other.model:
-            return True
-        return False
-
-    def set_model_sampler_cfg_function(self, sampler_cfg_function):
-        if len(inspect.signature(sampler_cfg_function).parameters) == 3:
-            self.model_options["sampler_cfg_function"] = lambda args: sampler_cfg_function(args["cond"], args["uncond"], args["cond_scale"]) #Old way
-        else:
-            self.model_options["sampler_cfg_function"] = sampler_cfg_function
-
-    def set_model_unet_function_wrapper(self, unet_wrapper_function):
-        self.model_options["model_function_wrapper"] = unet_wrapper_function
-
-    def set_model_patch(self, patch, name):
-        to = self.model_options["transformer_options"]
-        if "patches" not in to:
-            to["patches"] = {}
-        to["patches"][name] = to["patches"].get(name, []) + [patch]
-
-    def set_model_patch_replace(self, patch, name, block_name, number):
-        to = self.model_options["transformer_options"]
-        if "patches_replace" not in to:
-            to["patches_replace"] = {}
-        if name not in to["patches_replace"]:
-            to["patches_replace"][name] = {}
-        to["patches_replace"][name][(block_name, number)] = patch
-
-    def set_model_attn1_patch(self, patch):
-        self.set_model_patch(patch, "attn1_patch")
-
-    def set_model_attn2_patch(self, patch):
-        self.set_model_patch(patch, "attn2_patch")
-
-    def set_model_attn1_replace(self, patch, block_name, number):
-        self.set_model_patch_replace(patch, "attn1", block_name, number)
-
-    def set_model_attn2_replace(self, patch, block_name, number):
-        self.set_model_patch_replace(patch, "attn2", block_name, number)
-
-    def set_model_attn1_output_patch(self, patch):
-        self.set_model_patch(patch, "attn1_output_patch")
-
-    def set_model_attn2_output_patch(self, patch):
-        self.set_model_patch(patch, "attn2_output_patch")
-
-    def model_patches_to(self, device):
-        to = self.model_options["transformer_options"]
-        if "patches" in to:
-            patches = to["patches"]
-            for name in patches:
-                patch_list = patches[name]
-                for i in range(len(patch_list)):
-                    if hasattr(patch_list[i], "to"):
-                        patch_list[i] = patch_list[i].to(device)
-        if "patches_replace" in to:
-            patches = to["patches_replace"]
-            for name in patches:
-                patch_list = patches[name]
-                for k in patch_list:
-                    if hasattr(patch_list[k], "to"):
-                        patch_list[k] = patch_list[k].to(device)
-
-    def model_dtype(self):
-        if hasattr(self.model, "get_dtype"):
-            return self.model.get_dtype()
-
-    def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
-        p = set()
-        for k in patches:
-            if k in self.model_keys:
-                p.add(k)
-                current_patches = self.patches.get(k, [])
-                current_patches.append((strength_patch, patches[k], strength_model))
-                self.patches[k] = current_patches
-
-        return list(p)
-
-    def get_key_patches(self, filter_prefix=None):
-        model_sd = self.model_state_dict()
-        p = {}
-        for k in model_sd:
-            if filter_prefix is not None:
-                if not k.startswith(filter_prefix):
-                    continue
-            if k in self.patches:
-                p[k] = [model_sd[k]] + self.patches[k]
-            else:
-                p[k] = (model_sd[k],)
-        return p
-
-    def model_state_dict(self, filter_prefix=None):
-        sd = self.model.state_dict()
-        keys = list(sd.keys())
-        if filter_prefix is not None:
-            for k in keys:
-                if not k.startswith(filter_prefix):
-                    sd.pop(k)
-        return sd
-
-    def patch_model(self, device_to=None):
-        model_sd = self.model_state_dict()
-        for key in self.patches:
-            if key not in model_sd:
-                print("could not patch. key doesn't exist in model:", k)
-                continue
-
-            weight = model_sd[key]
-
-            if key not in self.backup:
-                self.backup[key] = weight.to(self.offload_device)
-
-            if device_to is not None:
-                temp_weight = weight.float().to(device_to, copy=True)
-            else:
-                temp_weight = weight.to(torch.float32, copy=True)
-            out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype)
-            set_attr(self.model, key, out_weight)
-            del temp_weight
-
-        if device_to is not None:
-            self.model.to(device_to)
-            self.current_device = device_to
-
-        return self.model
-
-    def calculate_weight(self, patches, weight, key):
-        for p in patches:
-            alpha = p[0]
-            v = p[1]
-            strength_model = p[2]
-
-            if strength_model != 1.0:
-                weight *= strength_model
-
-            if isinstance(v, list):
-                v = (self.calculate_weight(v[1:], v[0].clone(), key), )
-
-            if len(v) == 1:
-                w1 = v[0]
-                if alpha != 0.0:
-                    if w1.shape != weight.shape:
-                        print("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, w1.shape, weight.shape))
-                    else:
-                        weight += alpha * w1.type(weight.dtype).to(weight.device)
-            elif len(v) == 4: #lora/locon
-                mat1 = v[0].float().to(weight.device)
-                mat2 = v[1].float().to(weight.device)
-                if v[2] is not None:
-                    alpha *= v[2] / mat2.shape[0]
-                if v[3] is not None:
-                    #locon mid weights, hopefully the math is fine because I didn't properly test it
-                    mat3 = v[3].float().to(weight.device)
-                    final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
-                    mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
-                try:
-                    weight += (alpha * torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1))).reshape(weight.shape).type(weight.dtype)
-                except Exception as e:
-                    print("ERROR", key, e)
-            elif len(v) == 8: #lokr
-                w1 = v[0]
-                w2 = v[1]
-                w1_a = v[3]
-                w1_b = v[4]
-                w2_a = v[5]
-                w2_b = v[6]
-                t2 = v[7]
-                dim = None
-
-                if w1 is None:
-                    dim = w1_b.shape[0]
-                    w1 = torch.mm(w1_a.float(), w1_b.float())
-                else:
-                    w1 = w1.float().to(weight.device)
-
-                if w2 is None:
-                    dim = w2_b.shape[0]
-                    if t2 is None:
-                        w2 = torch.mm(w2_a.float().to(weight.device), w2_b.float().to(weight.device))
-                    else:
-                        w2 = torch.einsum('i j k l, j r, i p -> p r k l', t2.float().to(weight.device), w2_b.float().to(weight.device), w2_a.float().to(weight.device))
-                else:
-                    w2 = w2.float().to(weight.device)
-
-                if len(w2.shape) == 4:
-                    w1 = w1.unsqueeze(2).unsqueeze(2)
-                if v[2] is not None and dim is not None:
-                    alpha *= v[2] / dim
-
-                try:
-                    weight += alpha * torch.kron(w1, w2).reshape(weight.shape).type(weight.dtype)
-                except Exception as e:
-                    print("ERROR", key, e)
-            else: #loha
-                w1a = v[0]
-                w1b = v[1]
-                if v[2] is not None:
-                    alpha *= v[2] / w1b.shape[0]
-                w2a = v[3]
-                w2b = v[4]
-                if v[5] is not None: #cp decomposition
-                    t1 = v[5]
-                    t2 = v[6]
-                    m1 = torch.einsum('i j k l, j r, i p -> p r k l', t1.float().to(weight.device), w1b.float().to(weight.device), w1a.float().to(weight.device))
-                    m2 = torch.einsum('i j k l, j r, i p -> p r k l', t2.float().to(weight.device), w2b.float().to(weight.device), w2a.float().to(weight.device))
-                else:
-                    m1 = torch.mm(w1a.float().to(weight.device), w1b.float().to(weight.device))
-                    m2 = torch.mm(w2a.float().to(weight.device), w2b.float().to(weight.device))
-
-                try:
-                    weight += (alpha * m1 * m2).reshape(weight.shape).type(weight.dtype)
-                except Exception as e:
-                    print("ERROR", key, e)
-
-        return weight
-
-    def unpatch_model(self, device_to=None):
-        keys = list(self.backup.keys())
-
-        for k in keys:
-            set_attr(self.model, k, self.backup[k])
-
-        self.backup = {}
-
-        if device_to is not None:
-            self.model.to(device_to)
-            self.current_device = device_to
-
 
 def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
-    key_map = model_lora_keys_unet(model.model)
-    key_map = model_lora_keys_clip(clip.cond_stage_model, key_map)
-    loaded = load_lora(lora, key_map)
+    key_map = comfy.lora.model_lora_keys_unet(model.model)
+    key_map = comfy.lora.model_lora_keys_clip(clip.cond_stage_model, key_map)
+    loaded = comfy.lora.load_lora(lora, key_map)
     new_modelpatcher = model.clone()
     k = new_modelpatcher.add_patches(loaded, strength_model)
     new_clip = clip.clone()
@@ -544,7 +81,7 @@ class CLIP:
 
         load_device = model_management.text_encoder_device()
         offload_device = model_management.text_encoder_offload_device()
-        params['device'] = load_device
+        params['device'] = offload_device
         if model_management.should_use_fp16(load_device, prioritize_performance=False):
             params['dtype'] = torch.float16
         else:
@@ -553,7 +90,7 @@ class CLIP:
         self.cond_stage_model = clip(**(params))
 
         self.tokenizer = tokenizer(embedding_directory=embedding_directory)
-        self.patcher = ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
         self.layer_idx = None
 
     def clone(self):
@@ -564,9 +101,6 @@ class CLIP:
         n.layer_idx = self.layer_idx
         return n
 
-    def load_from_state_dict(self, sd):
-        self.cond_stage_model.load_sd(sd)
-
     def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
         return self.patcher.add_patches(patches, strength_patch, strength_model)
 
@@ -615,7 +149,7 @@ class VAE:
             self.first_stage_model = AutoencoderKL(**(config['params']))
         self.first_stage_model = self.first_stage_model.eval()
         if ckpt_path is not None:
-            sd = utils.load_torch_file(ckpt_path)
+            sd = comfy.utils.load_torch_file(ckpt_path)
             if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
                 sd = diffusers_convert.convert_vae_state_dict(sd)
             self.first_stage_model.load_state_dict(sd, strict=False)
@@ -628,29 +162,29 @@ class VAE:
         self.first_stage_model.to(self.vae_dtype)
 
     def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
-        steps = samples.shape[0] * utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap)
-        steps += samples.shape[0] * utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap)
-        steps += samples.shape[0] * utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap)
-        pbar = utils.ProgressBar(steps)
+        steps = samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap)
+        steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap)
+        steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap)
+        pbar = comfy.utils.ProgressBar(steps)
 
         decode_fn = lambda a: (self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)) + 1.0).float()
         output = torch.clamp((
-            (utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = 8, pbar = pbar) +
-            utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = 8, pbar = pbar) +
-             utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = 8, pbar = pbar))
+            (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = 8, pbar = pbar) +
+            comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = 8, pbar = pbar) +
+             comfy.utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = 8, pbar = pbar))
             / 3.0) / 2.0, min=0.0, max=1.0)
         return output
 
     def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
-        steps = pixel_samples.shape[0] * utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x, tile_y, overlap)
-        steps += pixel_samples.shape[0] * utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x // 2, tile_y * 2, overlap)
-        steps += pixel_samples.shape[0] * utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap)
-        pbar = utils.ProgressBar(steps)
+        steps = pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x, tile_y, overlap)
+        steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x // 2, tile_y * 2, overlap)
+        steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap)
+        pbar = comfy.utils.ProgressBar(steps)
 
         encode_fn = lambda a: self.first_stage_model.encode(2. * a.to(self.vae_dtype).to(self.device) - 1.).sample().float()
-        samples = utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
-        samples += utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
-        samples += utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
+        samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
+        samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
+        samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar)
         samples /= 3.0
         return samples
 
@@ -712,481 +246,6 @@ class VAE:
     def get_sd(self):
         return self.first_stage_model.state_dict()
 
-
-def broadcast_image_to(tensor, target_batch_size, batched_number):
-    current_batch_size = tensor.shape[0]
-    #print(current_batch_size, target_batch_size)
-    if current_batch_size == 1:
-        return tensor
-
-    per_batch = target_batch_size // batched_number
-    tensor = tensor[:per_batch]
-
-    if per_batch > tensor.shape[0]:
-        tensor = torch.cat([tensor] * (per_batch // tensor.shape[0]) + [tensor[:(per_batch % tensor.shape[0])]], dim=0)
-
-    current_batch_size = tensor.shape[0]
-    if current_batch_size == target_batch_size:
-        return tensor
-    else:
-        return torch.cat([tensor] * batched_number, dim=0)
-
-class ControlBase:
-    def __init__(self, device=None):
-        self.cond_hint_original = None
-        self.cond_hint = None
-        self.strength = 1.0
-        self.timestep_percent_range = (1.0, 0.0)
-        self.timestep_range = None
-
-        if device is None:
-            device = model_management.get_torch_device()
-        self.device = device
-        self.previous_controlnet = None
-        self.global_average_pooling = False
-
-    def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(1.0, 0.0)):
-        self.cond_hint_original = cond_hint
-        self.strength = strength
-        self.timestep_percent_range = timestep_percent_range
-        return self
-
-    def pre_run(self, model, percent_to_timestep_function):
-        self.timestep_range = (percent_to_timestep_function(self.timestep_percent_range[0]), percent_to_timestep_function(self.timestep_percent_range[1]))
-        if self.previous_controlnet is not None:
-            self.previous_controlnet.pre_run(model, percent_to_timestep_function)
-
-    def set_previous_controlnet(self, controlnet):
-        self.previous_controlnet = controlnet
-        return self
-
-    def cleanup(self):
-        if self.previous_controlnet is not None:
-            self.previous_controlnet.cleanup()
-        if self.cond_hint is not None:
-            del self.cond_hint
-            self.cond_hint = None
-        self.timestep_range = None
-
-    def get_models(self):
-        out = []
-        if self.previous_controlnet is not None:
-            out += self.previous_controlnet.get_models()
-        return out
-
-    def copy_to(self, c):
-        c.cond_hint_original = self.cond_hint_original
-        c.strength = self.strength
-        c.timestep_percent_range = self.timestep_percent_range
-
-    def inference_memory_requirements(self, dtype):
-        if self.previous_controlnet is not None:
-            return self.previous_controlnet.inference_memory_requirements(dtype)
-        return 0
-
-    def control_merge(self, control_input, control_output, control_prev, output_dtype):
-        out = {'input':[], 'middle':[], 'output': []}
-
-        if control_input is not None:
-            for i in range(len(control_input)):
-                key = 'input'
-                x = control_input[i]
-                if x is not None:
-                    x *= self.strength
-                    if x.dtype != output_dtype:
-                        x = x.to(output_dtype)
-                out[key].insert(0, x)
-
-        if control_output is not None:
-            for i in range(len(control_output)):
-                if i == (len(control_output) - 1):
-                    key = 'middle'
-                    index = 0
-                else:
-                    key = 'output'
-                    index = i
-                x = control_output[i]
-                if x is not None:
-                    if self.global_average_pooling:
-                        x = torch.mean(x, dim=(2, 3), keepdim=True).repeat(1, 1, x.shape[2], x.shape[3])
-
-                    x *= self.strength
-                    if x.dtype != output_dtype:
-                        x = x.to(output_dtype)
-
-                out[key].append(x)
-        if control_prev is not None:
-            for x in ['input', 'middle', 'output']:
-                o = out[x]
-                for i in range(len(control_prev[x])):
-                    prev_val = control_prev[x][i]
-                    if i >= len(o):
-                        o.append(prev_val)
-                    elif prev_val is not None:
-                        if o[i] is None:
-                            o[i] = prev_val
-                        else:
-                            o[i] += prev_val
-        return out
-
-class ControlNet(ControlBase):
-    def __init__(self, control_model, global_average_pooling=False, device=None):
-        super().__init__(device)
-        self.control_model = control_model
-        self.control_model_wrapped = ModelPatcher(self.control_model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
-        self.global_average_pooling = global_average_pooling
-
-    def get_control(self, x_noisy, t, cond, batched_number):
-        control_prev = None
-        if self.previous_controlnet is not None:
-            control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
-
-        if self.timestep_range is not None:
-            if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
-                if control_prev is not None:
-                    return control_prev
-                else:
-                    return {}
-
-        output_dtype = x_noisy.dtype
-        if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]:
-            if self.cond_hint is not None:
-                del self.cond_hint
-            self.cond_hint = None
-            self.cond_hint = utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * 8, x_noisy.shape[2] * 8, 'nearest-exact', "center").to(self.control_model.dtype).to(self.device)
-        if x_noisy.shape[0] != self.cond_hint.shape[0]:
-            self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
-
-
-        context = torch.cat(cond['c_crossattn'], 1)
-        y = cond.get('c_adm', None)
-        if y is not None:
-            y = y.to(self.control_model.dtype)
-        control = self.control_model(x=x_noisy.to(self.control_model.dtype), hint=self.cond_hint, timesteps=t, context=context.to(self.control_model.dtype), y=y)
-        return self.control_merge(None, control, control_prev, output_dtype)
-
-    def copy(self):
-        c = ControlNet(self.control_model, global_average_pooling=self.global_average_pooling)
-        self.copy_to(c)
-        return c
-
-    def get_models(self):
-        out = super().get_models()
-        out.append(self.control_model_wrapped)
-        return out
-
-class ControlLoraOps:
-    class Linear(torch.nn.Module):
-        def __init__(self, in_features: int, out_features: int, bias: bool = True,
-                    device=None, dtype=None) -> None:
-            factory_kwargs = {'device': device, 'dtype': dtype}
-            super().__init__()
-            self.in_features = in_features
-            self.out_features = out_features
-            self.weight = None
-            self.up = None
-            self.down = None
-            self.bias = None
-
-        def forward(self, input):
-            if self.up is not None:
-                return torch.nn.functional.linear(input, self.weight.to(input.device) + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), self.bias)
-            else:
-                return torch.nn.functional.linear(input, self.weight.to(input.device), self.bias)
-
-    class Conv2d(torch.nn.Module):
-        def __init__(
-            self,
-            in_channels,
-            out_channels,
-            kernel_size,
-            stride=1,
-            padding=0,
-            dilation=1,
-            groups=1,
-            bias=True,
-            padding_mode='zeros',
-            device=None,
-            dtype=None
-        ):
-            super().__init__()
-            self.in_channels = in_channels
-            self.out_channels = out_channels
-            self.kernel_size = kernel_size
-            self.stride = stride
-            self.padding = padding
-            self.dilation = dilation
-            self.transposed = False
-            self.output_padding = 0
-            self.groups = groups
-            self.padding_mode = padding_mode
-
-            self.weight = None
-            self.bias = None
-            self.up = None
-            self.down = None
-
-
-        def forward(self, input):
-            if self.up is not None:
-                return torch.nn.functional.conv2d(input, self.weight.to(input.device) + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), self.bias, self.stride, self.padding, self.dilation, self.groups)
-            else:
-                return torch.nn.functional.conv2d(input, self.weight.to(input.device), self.bias, self.stride, self.padding, self.dilation, self.groups)
-
-    def conv_nd(self, dims, *args, **kwargs):
-        if dims == 2:
-            return self.Conv2d(*args, **kwargs)
-        else:
-            raise ValueError(f"unsupported dimensions: {dims}")
-
-
-class ControlLora(ControlNet):
-    def __init__(self, control_weights, global_average_pooling=False, device=None):
-        ControlBase.__init__(self, device)
-        self.control_weights = control_weights
-        self.global_average_pooling = global_average_pooling
-
-    def pre_run(self, model, percent_to_timestep_function):
-        super().pre_run(model, percent_to_timestep_function)
-        controlnet_config = model.model_config.unet_config.copy()
-        controlnet_config.pop("out_channels")
-        controlnet_config["hint_channels"] = self.control_weights["input_hint_block.0.weight"].shape[1]
-        controlnet_config["operations"] = ControlLoraOps()
-        self.control_model = cldm.ControlNet(**controlnet_config)
-        dtype = model.get_dtype()
-        self.control_model.to(dtype)
-        self.control_model.to(model_management.get_torch_device())
-        diffusion_model = model.diffusion_model
-        sd = diffusion_model.state_dict()
-        cm = self.control_model.state_dict()
-
-        for k in sd:
-            weight = sd[k]
-            if weight.device == torch.device("meta"): #lowvram NOTE: this depends on the inner working of the accelerate library so it might break.
-                key_split = k.split('.')              # I have no idea why they don't just leave the weight there instead of using the meta device.
-                op = get_attr(diffusion_model, '.'.join(key_split[:-1]))
-                weight = op._hf_hook.weights_map[key_split[-1]]
-
-            try:
-                set_attr(self.control_model, k, weight)
-            except:
-                pass
-
-        for k in self.control_weights:
-            if k not in {"lora_controlnet"}:
-                set_attr(self.control_model, k, self.control_weights[k].to(dtype).to(model_management.get_torch_device()))
-
-    def copy(self):
-        c = ControlLora(self.control_weights, global_average_pooling=self.global_average_pooling)
-        self.copy_to(c)
-        return c
-
-    def cleanup(self):
-        del self.control_model
-        self.control_model = None
-        super().cleanup()
-
-    def get_models(self):
-        out = ControlBase.get_models(self)
-        return out
-
-    def inference_memory_requirements(self, dtype):
-        return utils.calculate_parameters(self.control_weights) * model_management.dtype_size(dtype) + ControlBase.inference_memory_requirements(self, dtype)
-
-def load_controlnet(ckpt_path, model=None):
-    controlnet_data = utils.load_torch_file(ckpt_path, safe_load=True)
-    if "lora_controlnet" in controlnet_data:
-        return ControlLora(controlnet_data)
-
-    controlnet_config = None
-    if "controlnet_cond_embedding.conv_in.weight" in controlnet_data: #diffusers format
-        use_fp16 = model_management.should_use_fp16()
-        controlnet_config = model_detection.unet_config_from_diffusers_unet(controlnet_data, use_fp16)
-        diffusers_keys = utils.unet_to_diffusers(controlnet_config)
-        diffusers_keys["controlnet_mid_block.weight"] = "middle_block_out.0.weight"
-        diffusers_keys["controlnet_mid_block.bias"] = "middle_block_out.0.bias"
-
-        count = 0
-        loop = True
-        while loop:
-            suffix = [".weight", ".bias"]
-            for s in suffix:
-                k_in = "controlnet_down_blocks.{}{}".format(count, s)
-                k_out = "zero_convs.{}.0{}".format(count, s)
-                if k_in not in controlnet_data:
-                    loop = False
-                    break
-                diffusers_keys[k_in] = k_out
-            count += 1
-
-        count = 0
-        loop = True
-        while loop:
-            suffix = [".weight", ".bias"]
-            for s in suffix:
-                if count == 0:
-                    k_in = "controlnet_cond_embedding.conv_in{}".format(s)
-                else:
-                    k_in = "controlnet_cond_embedding.blocks.{}{}".format(count - 1, s)
-                k_out = "input_hint_block.{}{}".format(count * 2, s)
-                if k_in not in controlnet_data:
-                    k_in = "controlnet_cond_embedding.conv_out{}".format(s)
-                    loop = False
-                diffusers_keys[k_in] = k_out
-            count += 1
-
-        new_sd = {}
-        for k in diffusers_keys:
-            if k in controlnet_data:
-                new_sd[diffusers_keys[k]] = controlnet_data.pop(k)
-
-        leftover_keys = controlnet_data.keys()
-        if len(leftover_keys) > 0:
-            print("leftover keys:", leftover_keys)
-        controlnet_data = new_sd
-
-    pth_key = 'control_model.zero_convs.0.0.weight'
-    pth = False
-    key = 'zero_convs.0.0.weight'
-    if pth_key in controlnet_data:
-        pth = True
-        key = pth_key
-        prefix = "control_model."
-    elif key in controlnet_data:
-        prefix = ""
-    else:
-        net = load_t2i_adapter(controlnet_data)
-        if net is None:
-            print("error checkpoint does not contain controlnet or t2i adapter data", ckpt_path)
-        return net
-
-    if controlnet_config is None:
-        use_fp16 = model_management.should_use_fp16()
-        controlnet_config = model_detection.model_config_from_unet(controlnet_data, prefix, use_fp16).unet_config
-    controlnet_config.pop("out_channels")
-    controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
-    control_model = cldm.ControlNet(**controlnet_config)
-
-    if pth:
-        if 'difference' in controlnet_data:
-            if model is not None:
-                model_management.load_models_gpu([model])
-                model_sd = model.model_state_dict()
-                for x in controlnet_data:
-                    c_m = "control_model."
-                    if x.startswith(c_m):
-                        sd_key = "diffusion_model.{}".format(x[len(c_m):])
-                        if sd_key in model_sd:
-                            cd = controlnet_data[x]
-                            cd += model_sd[sd_key].type(cd.dtype).to(cd.device)
-            else:
-                print("WARNING: Loaded a diff controlnet without a model. It will very likely not work.")
-
-        class WeightsLoader(torch.nn.Module):
-            pass
-        w = WeightsLoader()
-        w.control_model = control_model
-        missing, unexpected = w.load_state_dict(controlnet_data, strict=False)
-    else:
-        missing, unexpected = control_model.load_state_dict(controlnet_data, strict=False)
-    print(missing, unexpected)
-
-    if use_fp16:
-        control_model = control_model.half()
-
-    global_average_pooling = False
-    if ckpt_path.endswith("_shuffle.pth") or ckpt_path.endswith("_shuffle.safetensors") or ckpt_path.endswith("_shuffle_fp16.safetensors"): #TODO: smarter way of enabling global_average_pooling
-        global_average_pooling = True
-
-    control = ControlNet(control_model, global_average_pooling=global_average_pooling)
-    return control
-
-class T2IAdapter(ControlBase):
-    def __init__(self, t2i_model, channels_in, device=None):
-        super().__init__(device)
-        self.t2i_model = t2i_model
-        self.channels_in = channels_in
-        self.control_input = None
-
-    def scale_image_to(self, width, height):
-        unshuffle_amount = self.t2i_model.unshuffle_amount
-        width = math.ceil(width / unshuffle_amount) * unshuffle_amount
-        height = math.ceil(height / unshuffle_amount) * unshuffle_amount
-        return width, height
-
-    def get_control(self, x_noisy, t, cond, batched_number):
-        control_prev = None
-        if self.previous_controlnet is not None:
-            control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
-
-        if self.timestep_range is not None:
-            if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
-                if control_prev is not None:
-                    return control_prev
-                else:
-                    return {}
-
-        if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]:
-            if self.cond_hint is not None:
-                del self.cond_hint
-            self.control_input = None
-            self.cond_hint = None
-            width, height = self.scale_image_to(x_noisy.shape[3] * 8, x_noisy.shape[2] * 8)
-            self.cond_hint = utils.common_upscale(self.cond_hint_original, width, height, 'nearest-exact', "center").float().to(self.device)
-            if self.channels_in == 1 and self.cond_hint.shape[1] > 1:
-                self.cond_hint = torch.mean(self.cond_hint, 1, keepdim=True)
-        if x_noisy.shape[0] != self.cond_hint.shape[0]:
-            self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
-        if self.control_input is None:
-            self.t2i_model.to(x_noisy.dtype)
-            self.t2i_model.to(self.device)
-            self.control_input = self.t2i_model(self.cond_hint.to(x_noisy.dtype))
-            self.t2i_model.cpu()
-
-        control_input = list(map(lambda a: None if a is None else a.clone(), self.control_input))
-        mid = None
-        if self.t2i_model.xl == True:
-            mid = control_input[-1:]
-            control_input = control_input[:-1]
-        return self.control_merge(control_input, mid, control_prev, x_noisy.dtype)
-
-    def copy(self):
-        c = T2IAdapter(self.t2i_model, self.channels_in)
-        self.copy_to(c)
-        return c
-
-def load_t2i_adapter(t2i_data):
-    keys = t2i_data.keys()
-    if 'adapter' in keys:
-        t2i_data = t2i_data['adapter']
-        keys = t2i_data.keys()
-    if "body.0.in_conv.weight" in keys:
-        cin = t2i_data['body.0.in_conv.weight'].shape[1]
-        model_ad = adapter.Adapter_light(cin=cin, channels=[320, 640, 1280, 1280], nums_rb=4)
-    elif 'conv_in.weight' in keys:
-        cin = t2i_data['conv_in.weight'].shape[1]
-        channel = t2i_data['conv_in.weight'].shape[0]
-        ksize = t2i_data['body.0.block2.weight'].shape[2]
-        use_conv = False
-        down_opts = list(filter(lambda a: a.endswith("down_opt.op.weight"), keys))
-        if len(down_opts) > 0:
-            use_conv = True
-        xl = False
-        if cin == 256:
-            xl = True
-        model_ad = adapter.Adapter(cin=cin, channels=[channel, channel*2, channel*4, channel*4][:4], nums_rb=2, ksize=ksize, sk=True, use_conv=use_conv, xl=xl)
-    else:
-        return None
-    missing, unexpected = model_ad.load_state_dict(t2i_data)
-    if len(missing) > 0:
-        print("t2i missing", missing)
-
-    if len(unexpected) > 0:
-        print("t2i unexpected", unexpected)
-
-    return T2IAdapter(model_ad, model_ad.input_channels)
-
-
 class StyleModel:
     def __init__(self, model, device="cpu"):
         self.model = model
@@ -1196,10 +255,10 @@ class StyleModel:
 
 
 def load_style_model(ckpt_path):
-    model_data = utils.load_torch_file(ckpt_path, safe_load=True)
+    model_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
     keys = model_data.keys()
     if "style_embedding" in keys:
-        model = adapter.StyleAdapter(width=1024, context_dim=768, num_head=8, n_layes=3, num_token=8)
+        model = comfy.t2i_adapter.adapter.StyleAdapter(width=1024, context_dim=768, num_head=8, n_layes=3, num_token=8)
     else:
         raise Exception("invalid style model {}".format(ckpt_path))
     model.load_state_dict(model_data)
@@ -1209,14 +268,14 @@ def load_style_model(ckpt_path):
 def load_clip(ckpt_paths, embedding_directory=None):
     clip_data = []
     for p in ckpt_paths:
-        clip_data.append(utils.load_torch_file(p, safe_load=True))
+        clip_data.append(comfy.utils.load_torch_file(p, safe_load=True))
 
     class EmptyClass:
         pass
 
     for i in range(len(clip_data)):
         if "transformer.resblocks.0.ln_1.weight" in clip_data[i]:
-            clip_data[i] = utils.transformers_convert(clip_data[i], "", "text_model.", 32)
+            clip_data[i] = comfy.utils.transformers_convert(clip_data[i], "", "text_model.", 32)
 
     clip_target = EmptyClass()
     clip_target.params = {}
@@ -1245,11 +304,11 @@ def load_clip(ckpt_paths, embedding_directory=None):
     return clip
 
 def load_gligen(ckpt_path):
-    data = utils.load_torch_file(ckpt_path, safe_load=True)
+    data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
     model = gligen.load_gligen(data)
     if model_management.should_use_fp16():
         model = model.half()
-    return ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
+    return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
 
 def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_clip=True, embedding_directory=None, state_dict=None, config=None):
     #TODO: this function is a mess and should be removed eventually
@@ -1285,23 +344,25 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl
         pass
 
     if state_dict is None:
-        state_dict = utils.load_torch_file(ckpt_path)
+        state_dict = comfy.utils.load_torch_file(ckpt_path)
 
     class EmptyClass:
         pass
 
-    model_config = EmptyClass()
-    model_config.unet_config = unet_config
+    model_config = comfy.supported_models_base.BASE({})
+
     from . import latent_formats
     model_config.latent_format = latent_formats.SD15(scale_factor=scale_factor)
+    model_config.unet_config = unet_config
 
-    if config['model']["target"].endswith("LatentInpaintDiffusion"):
-        model = model_base.SDInpaint(model_config, model_type=model_type)
-    elif config['model']["target"].endswith("ImageEmbeddingConditionedLatentDiffusion"):
+    if config['model']["target"].endswith("ImageEmbeddingConditionedLatentDiffusion"):
         model = model_base.SD21UNCLIP(model_config, noise_aug_config["params"], model_type=model_type)
     else:
         model = model_base.BaseModel(model_config, model_type=model_type)
 
+    if config['model']["target"].endswith("LatentInpaintDiffusion"):
+        model.set_inpaint()
+
     if fp16:
         model = model.half()
 
@@ -1329,10 +390,10 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl
         w.cond_stage_model = clip.cond_stage_model
         load_clip_weights(w, state_dict)
 
-    return (ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device), clip, vae)
+    return (comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device), clip, vae)
 
 def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None):
-    sd = utils.load_torch_file(ckpt_path)
+    sd = comfy.utils.load_torch_file(ckpt_path)
     sd_keys = sd.keys()
     clip = None
     clipvision = None
@@ -1340,7 +401,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
     model = None
     clip_target = None
 
-    parameters = utils.calculate_parameters(sd, "model.diffusion_model.")
+    parameters = comfy.utils.calculate_parameters(sd, "model.diffusion_model.")
     fp16 = model_management.should_use_fp16(model_params=parameters)
 
     class WeightsLoader(torch.nn.Module):
@@ -1381,7 +442,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
     if len(left_over) > 0:
         print("left over keys:", left_over)
 
-    model_patcher = ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device(), current_device=inital_load_device)
+    model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device(), current_device=inital_load_device)
     if inital_load_device != torch.device("cpu"):
         print("loaded straight to GPU")
         model_management.load_model_gpu(model_patcher)
@@ -1390,8 +451,8 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
 
 
 def load_unet(unet_path): #load unet in diffusers format
-    sd = utils.load_torch_file(unet_path)
-    parameters = utils.calculate_parameters(sd)
+    sd = comfy.utils.load_torch_file(unet_path)
+    parameters = comfy.utils.calculate_parameters(sd)
     fp16 = model_management.should_use_fp16(model_params=parameters)
 
     model_config = model_detection.model_config_from_diffusers_unet(sd, fp16)
@@ -1399,7 +460,7 @@ def load_unet(unet_path): #load unet in diffusers format
         print("ERROR UNSUPPORTED UNET", unet_path)
         return None
 
-    diffusers_keys = utils.unet_to_diffusers(model_config.unet_config)
+    diffusers_keys = comfy.utils.unet_to_diffusers(model_config.unet_config)
 
     new_sd = {}
     for k in diffusers_keys:
@@ -1411,9 +472,9 @@ def load_unet(unet_path): #load unet in diffusers format
     model = model_config.get_model(new_sd, "")
     model = model.to(offload_device)
     model.load_model_weights(new_sd, "")
-    return ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device)
+    return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device)
 
 def save_checkpoint(output_path, model, clip, vae, metadata=None):
     model_management.load_models_gpu([model, clip.load_model()])
     sd = model.model.state_dict_for_saving(clip.get_sd(), vae.get_sd())
-    utils.save_torch_file(sd, output_path, metadata=metadata)
+    comfy.utils.save_torch_file(sd, output_path, metadata=metadata)
diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py
index 4616ca4e..477d5c30 100644
--- a/comfy/sd1_clip.py
+++ b/comfy/sd1_clip.py
@@ -66,7 +66,9 @@ class SD1ClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.layer = layer
         self.layer_idx = None
         self.empty_tokens = [[49406] + [49407] * 76]
-        self.text_projection = None
+        self.text_projection = torch.nn.Parameter(torch.eye(self.transformer.get_input_embeddings().weight.shape[1]))
+        self.logit_scale = torch.nn.Parameter(torch.tensor(4.6055))
+
         self.layer_norm_hidden_state = True
         if layer == "hidden":
             assert layer_idx is not None
@@ -163,6 +165,10 @@ class SD1ClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         return self(tokens)
 
     def load_sd(self, sd):
+        if "text_projection" in sd:
+            self.text_projection[:] = sd.pop("text_projection")
+        if "text_projection.weight" in sd:
+            self.text_projection[:] = sd.pop("text_projection.weight").transpose(0, 1)
         return self.transformer.load_state_dict(sd, strict=False)
 
 def parse_parentheses(string):
diff --git a/comfy/sd2_clip_config.json b/comfy/sd2_clip_config.json
index ace6ef00..85cec832 100644
--- a/comfy/sd2_clip_config.json
+++ b/comfy/sd2_clip_config.json
@@ -17,7 +17,7 @@
   "num_attention_heads": 16,
   "num_hidden_layers": 24,
   "pad_token_id": 1,
-  "projection_dim": 512,
+  "projection_dim": 1024,
   "torch_dtype": "float32",
   "vocab_size": 49408
 }
diff --git a/comfy/sdxl_clip.py b/comfy/sdxl_clip.py
index d05c0a9b..e3ac2ee0 100644
--- a/comfy/sdxl_clip.py
+++ b/comfy/sdxl_clip.py
@@ -11,15 +11,9 @@ class SDXLClipG(sd1_clip.SD1ClipModel):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, textmodel_path=textmodel_path, dtype=dtype)
         self.empty_tokens = [[49406] + [49407] + [0] * 75]
-        self.text_projection = torch.nn.Parameter(torch.empty(1280, 1280))
-        self.logit_scale = torch.nn.Parameter(torch.tensor(4.6055))
         self.layer_norm_hidden_state = False
 
     def load_sd(self, sd):
-        if "text_projection" in sd:
-            self.text_projection[:] = sd.pop("text_projection")
-        if "text_projection.weight" in sd:
-            self.text_projection[:] = sd.pop("text_projection.weight").transpose(0, 1)
         return super().load_sd(sd)
 
 class SDXLClipGTokenizer(sd1_clip.SD1Tokenizer):
diff --git a/comfy/supported_models.py b/comfy/supported_models.py
index 95fc8f3f..bb8ae214 100644
--- a/comfy/supported_models.py
+++ b/comfy/supported_models.py
@@ -68,7 +68,7 @@ class SD20(supported_models_base.BASE):
     def process_clip_state_dict_for_saving(self, state_dict):
         replace_prefix = {}
         replace_prefix[""] = "cond_stage_model.model."
-        state_dict = supported_models_base.state_dict_prefix_replace(state_dict, replace_prefix)
+        state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix)
         state_dict = diffusers_convert.convert_text_enc_state_dict_v20(state_dict)
         return state_dict
 
@@ -120,7 +120,7 @@ class SDXLRefiner(supported_models_base.BASE):
         keys_to_replace["conditioner.embedders.0.model.text_projection"] = "cond_stage_model.clip_g.text_projection"
         keys_to_replace["conditioner.embedders.0.model.logit_scale"] = "cond_stage_model.clip_g.logit_scale"
 
-        state_dict = supported_models_base.state_dict_key_replace(state_dict, keys_to_replace)
+        state_dict = utils.state_dict_key_replace(state_dict, keys_to_replace)
         return state_dict
 
     def process_clip_state_dict_for_saving(self, state_dict):
@@ -129,7 +129,7 @@ class SDXLRefiner(supported_models_base.BASE):
         if "clip_g.transformer.text_model.embeddings.position_ids" in state_dict_g:
             state_dict_g.pop("clip_g.transformer.text_model.embeddings.position_ids")
         replace_prefix["clip_g"] = "conditioner.embedders.0.model"
-        state_dict_g = supported_models_base.state_dict_prefix_replace(state_dict_g, replace_prefix)
+        state_dict_g = utils.state_dict_prefix_replace(state_dict_g, replace_prefix)
         return state_dict_g
 
     def clip_target(self):
@@ -153,7 +153,10 @@ class SDXL(supported_models_base.BASE):
             return model_base.ModelType.EPS
 
     def get_model(self, state_dict, prefix="", device=None):
-        return model_base.SDXL(self, model_type=self.model_type(state_dict, prefix), device=device)
+        out = model_base.SDXL(self, model_type=self.model_type(state_dict, prefix), device=device)
+        if self.inpaint_model():
+            out.set_inpaint()
+        return out
 
     def process_clip_state_dict(self, state_dict):
         keys_to_replace = {}
@@ -164,8 +167,8 @@ class SDXL(supported_models_base.BASE):
         keys_to_replace["conditioner.embedders.1.model.text_projection"] = "cond_stage_model.clip_g.text_projection"
         keys_to_replace["conditioner.embedders.1.model.logit_scale"] = "cond_stage_model.clip_g.logit_scale"
 
-        state_dict = supported_models_base.state_dict_prefix_replace(state_dict, replace_prefix)
-        state_dict = supported_models_base.state_dict_key_replace(state_dict, keys_to_replace)
+        state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix)
+        state_dict = utils.state_dict_key_replace(state_dict, keys_to_replace)
         return state_dict
 
     def process_clip_state_dict_for_saving(self, state_dict):
@@ -180,7 +183,7 @@ class SDXL(supported_models_base.BASE):
 
         replace_prefix["clip_g"] = "conditioner.embedders.1.model"
         replace_prefix["clip_l"] = "conditioner.embedders.0"
-        state_dict_g = supported_models_base.state_dict_prefix_replace(state_dict_g, replace_prefix)
+        state_dict_g = utils.state_dict_prefix_replace(state_dict_g, replace_prefix)
         return state_dict_g
 
     def clip_target(self):
diff --git a/comfy/supported_models_base.py b/comfy/supported_models_base.py
index d0088bbd..88a1d7fd 100644
--- a/comfy/supported_models_base.py
+++ b/comfy/supported_models_base.py
@@ -1,21 +1,7 @@
 import torch
 from . import model_base
 from . import utils
-
-
-def state_dict_key_replace(state_dict, keys_to_replace):
-    for x in keys_to_replace:
-        if x in state_dict:
-            state_dict[keys_to_replace[x]] = state_dict.pop(x)
-    return state_dict
-
-def state_dict_prefix_replace(state_dict, replace_prefix):
-    for rp in replace_prefix:
-        replace = list(map(lambda a: (a, "{}{}".format(replace_prefix[rp], a[len(rp):])), filter(lambda a: a.startswith(rp), state_dict.keys())))
-        for x in replace:
-            state_dict[x[1]] = state_dict.pop(x[0])
-    return state_dict
-
+from . import latent_formats
 
 class ClipTarget:
     def __init__(self, tokenizer, clip):
@@ -33,6 +19,8 @@ class BASE:
     clip_prefix = []
     clip_vision_prefix = None
     noise_aug_config = None
+    beta_schedule = "linear"
+    latent_format = latent_formats.LatentFormat
 
     @classmethod
     def matches(s, unet_config):
@@ -54,25 +42,26 @@ class BASE:
             self.unet_config[x] = self.unet_extra_config[x]
 
     def get_model(self, state_dict, prefix="", device=None):
-        if self.inpaint_model():
-            return model_base.SDInpaint(self, model_type=self.model_type(state_dict, prefix), device=device)
-        elif self.noise_aug_config is not None:
-            return model_base.SD21UNCLIP(self, self.noise_aug_config, model_type=self.model_type(state_dict, prefix), device=device)
+        if self.noise_aug_config is not None:
+            out = model_base.SD21UNCLIP(self, self.noise_aug_config, model_type=self.model_type(state_dict, prefix), device=device)
         else:
-            return model_base.BaseModel(self, model_type=self.model_type(state_dict, prefix), device=device)
+            out = model_base.BaseModel(self, model_type=self.model_type(state_dict, prefix), device=device)
+        if self.inpaint_model():
+            out.set_inpaint()
+        return out
 
     def process_clip_state_dict(self, state_dict):
         return state_dict
 
     def process_clip_state_dict_for_saving(self, state_dict):
         replace_prefix = {"": "cond_stage_model."}
-        return state_dict_prefix_replace(state_dict, replace_prefix)
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
 
     def process_unet_state_dict_for_saving(self, state_dict):
         replace_prefix = {"": "model.diffusion_model."}
-        return state_dict_prefix_replace(state_dict, replace_prefix)
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
 
     def process_vae_state_dict_for_saving(self, state_dict):
         replace_prefix = {"": "first_stage_model."}
-        return state_dict_prefix_replace(state_dict, replace_prefix)
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
 
diff --git a/comfy/utils.py b/comfy/utils.py
index e69125ab..3ed32e37 100644
--- a/comfy/utils.py
+++ b/comfy/utils.py
@@ -39,6 +39,20 @@ def calculate_parameters(sd, prefix=""):
             params += sd[k].nelement()
     return params
 
+def state_dict_key_replace(state_dict, keys_to_replace):
+    for x in keys_to_replace:
+        if x in state_dict:
+            state_dict[keys_to_replace[x]] = state_dict.pop(x)
+    return state_dict
+
+def state_dict_prefix_replace(state_dict, replace_prefix):
+    for rp in replace_prefix:
+        replace = list(map(lambda a: (a, "{}{}".format(replace_prefix[rp], a[len(rp):])), filter(lambda a: a.startswith(rp), state_dict.keys())))
+        for x in replace:
+            state_dict[x[1]] = state_dict.pop(x[0])
+    return state_dict
+
+
 def transformers_convert(sd, prefix_from, prefix_to, number):
     keys_to_replace = {
         "{}positional_embedding": "{}embeddings.position_embedding.weight",
@@ -223,6 +237,13 @@ def unet_to_diffusers(unet_config):
 
     return diffusers_unet_map
 
+def repeat_to_batch_size(tensor, batch_size):
+    if tensor.shape[0] > batch_size:
+        return tensor[:batch_size]
+    elif tensor.shape[0] < batch_size:
+        return tensor.repeat([math.ceil(batch_size / tensor.shape[0])] + [1] * (len(tensor.shape) - 1))[:batch_size]
+    return tensor
+
 def convert_sd_to(state_dict, dtype):
     keys = list(state_dict.keys())
     for k in keys:
@@ -237,6 +258,20 @@ def safetensors_header(safetensors_path, max_size=100*1024*1024):
             return None
         return f.read(length_of_header)
 
+def set_attr(obj, attr, value):
+    attrs = attr.split(".")
+    for name in attrs[:-1]:
+        obj = getattr(obj, name)
+    prev = getattr(obj, attrs[-1])
+    setattr(obj, attrs[-1], torch.nn.Parameter(value))
+    del prev
+
+def get_attr(obj, attr):
+    attrs = attr.split(".")
+    for name in attrs:
+        obj = getattr(obj, name)
+    return obj
+
 def bislerp(samples, width, height):
     def slerp(b1, b2, r):
         '''slerps batches b1, b2 according to ratio r, batches should be flat e.g. NxC'''
diff --git a/comfy_extras/chainner_models/architecture/DAT.py b/comfy_extras/chainner_models/architecture/DAT.py
new file mode 100644
index 00000000..0bcc26ef
--- /dev/null
+++ b/comfy_extras/chainner_models/architecture/DAT.py
@@ -0,0 +1,1182 @@
+# pylint: skip-file
+import math
+import re
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as checkpoint
+from einops import rearrange
+from einops.layers.torch import Rearrange
+from torch import Tensor
+from torch.nn import functional as F
+
+from .timm.drop import DropPath
+from .timm.weight_init import trunc_normal_
+
+
+def img2windows(img, H_sp, W_sp):
+    """
+    Input: Image (B, C, H, W)
+    Output: Window Partition (B', N, C)
+    """
+    B, C, H, W = img.shape
+    img_reshape = img.view(B, C, H // H_sp, H_sp, W // W_sp, W_sp)
+    img_perm = (
+        img_reshape.permute(0, 2, 4, 3, 5, 1).contiguous().reshape(-1, H_sp * W_sp, C)
+    )
+    return img_perm
+
+
+def windows2img(img_splits_hw, H_sp, W_sp, H, W):
+    """
+    Input: Window Partition (B', N, C)
+    Output: Image (B, H, W, C)
+    """
+    B = int(img_splits_hw.shape[0] / (H * W / H_sp / W_sp))
+
+    img = img_splits_hw.view(B, H // H_sp, W // W_sp, H_sp, W_sp, -1)
+    img = img.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return img
+
+
+class SpatialGate(nn.Module):
+    """Spatial-Gate.
+    Args:
+        dim (int): Half of input channels.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.conv = nn.Conv2d(
+            dim, dim, kernel_size=3, stride=1, padding=1, groups=dim
+        )  # DW Conv
+
+    def forward(self, x, H, W):
+        # Split
+        x1, x2 = x.chunk(2, dim=-1)
+        B, N, C = x.shape
+        x2 = (
+            self.conv(self.norm(x2).transpose(1, 2).contiguous().view(B, C // 2, H, W))
+            .flatten(2)
+            .transpose(-1, -2)
+            .contiguous()
+        )
+
+        return x1 * x2
+
+
+class SGFN(nn.Module):
+    """Spatial-Gate Feed-Forward Network.
+    Args:
+        in_features (int): Number of input channels.
+        hidden_features (int | None): Number of hidden channels. Default: None
+        out_features (int | None): Number of output channels. Default: None
+        act_layer (nn.Module): Activation layer. Default: nn.GELU
+        drop (float): Dropout rate. Default: 0.0
+    """
+
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        drop=0.0,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.sg = SpatialGate(hidden_features // 2)
+        self.fc2 = nn.Linear(hidden_features // 2, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x, H, W):
+        """
+        Input: x: (B, H*W, C), H, W
+        Output: x: (B, H*W, C)
+        """
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+
+        x = self.sg(x, H, W)
+        x = self.drop(x)
+
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class DynamicPosBias(nn.Module):
+    # The implementation builds on Crossformer code https://github.com/cheerss/CrossFormer/blob/main/models/crossformer.py
+    """Dynamic Relative Position Bias.
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        residual (bool):  If True, use residual strage to connect conv.
+    """
+
+    def __init__(self, dim, num_heads, residual):
+        super().__init__()
+        self.residual = residual
+        self.num_heads = num_heads
+        self.pos_dim = dim // 4
+        self.pos_proj = nn.Linear(2, self.pos_dim)
+        self.pos1 = nn.Sequential(
+            nn.LayerNorm(self.pos_dim),
+            nn.ReLU(inplace=True),
+            nn.Linear(self.pos_dim, self.pos_dim),
+        )
+        self.pos2 = nn.Sequential(
+            nn.LayerNorm(self.pos_dim),
+            nn.ReLU(inplace=True),
+            nn.Linear(self.pos_dim, self.pos_dim),
+        )
+        self.pos3 = nn.Sequential(
+            nn.LayerNorm(self.pos_dim),
+            nn.ReLU(inplace=True),
+            nn.Linear(self.pos_dim, self.num_heads),
+        )
+
+    def forward(self, biases):
+        if self.residual:
+            pos = self.pos_proj(biases)  # 2Gh-1 * 2Gw-1, heads
+            pos = pos + self.pos1(pos)
+            pos = pos + self.pos2(pos)
+            pos = self.pos3(pos)
+        else:
+            pos = self.pos3(self.pos2(self.pos1(self.pos_proj(biases))))
+        return pos
+
+
+class Spatial_Attention(nn.Module):
+    """Spatial Window Self-Attention.
+    It supports rectangle window (containing square window).
+    Args:
+        dim (int): Number of input channels.
+        idx (int): The indentix of window. (0/1)
+        split_size (tuple(int)): Height and Width of spatial window.
+        dim_out (int | None): The dimension of the attention output. Default: None
+        num_heads (int): Number of attention heads. Default: 6
+        attn_drop (float): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float): Dropout ratio of output. Default: 0.0
+        qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set
+        position_bias (bool): The dynamic relative position bias. Default: True
+    """
+
+    def __init__(
+        self,
+        dim,
+        idx,
+        split_size=[8, 8],
+        dim_out=None,
+        num_heads=6,
+        attn_drop=0.0,
+        proj_drop=0.0,
+        qk_scale=None,
+        position_bias=True,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.dim_out = dim_out or dim
+        self.split_size = split_size
+        self.num_heads = num_heads
+        self.idx = idx
+        self.position_bias = position_bias
+
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        if idx == 0:
+            H_sp, W_sp = self.split_size[0], self.split_size[1]
+        elif idx == 1:
+            W_sp, H_sp = self.split_size[0], self.split_size[1]
+        else:
+            print("ERROR MODE", idx)
+            exit(0)
+        self.H_sp = H_sp
+        self.W_sp = W_sp
+
+        if self.position_bias:
+            self.pos = DynamicPosBias(self.dim // 4, self.num_heads, residual=False)
+            # generate mother-set
+            position_bias_h = torch.arange(1 - self.H_sp, self.H_sp)
+            position_bias_w = torch.arange(1 - self.W_sp, self.W_sp)
+            biases = torch.stack(torch.meshgrid([position_bias_h, position_bias_w]))
+            biases = biases.flatten(1).transpose(0, 1).contiguous().float()
+            self.register_buffer("rpe_biases", biases)
+
+            # get pair-wise relative position index for each token inside the window
+            coords_h = torch.arange(self.H_sp)
+            coords_w = torch.arange(self.W_sp)
+            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))
+            coords_flatten = torch.flatten(coords, 1)
+            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
+            relative_coords = relative_coords.permute(1, 2, 0).contiguous()
+            relative_coords[:, :, 0] += self.H_sp - 1
+            relative_coords[:, :, 1] += self.W_sp - 1
+            relative_coords[:, :, 0] *= 2 * self.W_sp - 1
+            relative_position_index = relative_coords.sum(-1)
+            self.register_buffer("relative_position_index", relative_position_index)
+
+        self.attn_drop = nn.Dropout(attn_drop)
+
+    def im2win(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(-2, -1).contiguous().view(B, C, H, W)
+        x = img2windows(x, self.H_sp, self.W_sp)
+        x = (
+            x.reshape(-1, self.H_sp * self.W_sp, self.num_heads, C // self.num_heads)
+            .permute(0, 2, 1, 3)
+            .contiguous()
+        )
+        return x
+
+    def forward(self, qkv, H, W, mask=None):
+        """
+        Input: qkv: (B, 3*L, C), H, W, mask: (B, N, N), N is the window size
+        Output: x (B, H, W, C)
+        """
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        B, L, C = q.shape
+        assert L == H * W, "flatten img_tokens has wrong size"
+
+        # partition the q,k,v, image to window
+        q = self.im2win(q, H, W)
+        k = self.im2win(k, H, W)
+        v = self.im2win(v, H, W)
+
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)  # B head N C @ B head C N --> B head N N
+
+        # calculate drpe
+        if self.position_bias:
+            pos = self.pos(self.rpe_biases)
+            # select position bias
+            relative_position_bias = pos[self.relative_position_index.view(-1)].view(
+                self.H_sp * self.W_sp, self.H_sp * self.W_sp, -1
+            )
+            relative_position_bias = relative_position_bias.permute(
+                2, 0, 1
+            ).contiguous()
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+        N = attn.shape[3]
+
+        # use mask for shift window
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(
+                0
+            )
+            attn = attn.view(-1, self.num_heads, N, N)
+
+        attn = nn.functional.softmax(attn, dim=-1, dtype=attn.dtype)
+        attn = self.attn_drop(attn)
+
+        x = attn @ v
+        x = x.transpose(1, 2).reshape(
+            -1, self.H_sp * self.W_sp, C
+        )  # B head N N @ B head N C
+
+        # merge the window, window to image
+        x = windows2img(x, self.H_sp, self.W_sp, H, W)  # B H' W' C
+
+        return x
+
+
+class Adaptive_Spatial_Attention(nn.Module):
+    # The implementation builds on CAT code https://github.com/Zhengchen1999/CAT
+    """Adaptive Spatial Self-Attention
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads. Default: 6
+        split_size (tuple(int)): Height and Width of spatial window.
+        shift_size (tuple(int)): Shift size for spatial window.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float): Dropout rate. Default: 0.0
+        attn_drop (float): Attention dropout rate. Default: 0.0
+        rg_idx (int): The indentix of Residual Group (RG)
+        b_idx (int): The indentix of Block in each RG
+    """
+
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        reso=64,
+        split_size=[8, 8],
+        shift_size=[1, 2],
+        qkv_bias=False,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        rg_idx=0,
+        b_idx=0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.split_size = split_size
+        self.shift_size = shift_size
+        self.b_idx = b_idx
+        self.rg_idx = rg_idx
+        self.patches_resolution = reso
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+
+        assert (
+            0 <= self.shift_size[0] < self.split_size[0]
+        ), "shift_size must in 0-split_size0"
+        assert (
+            0 <= self.shift_size[1] < self.split_size[1]
+        ), "shift_size must in 0-split_size1"
+
+        self.branch_num = 2
+
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(drop)
+
+        self.attns = nn.ModuleList(
+            [
+                Spatial_Attention(
+                    dim // 2,
+                    idx=i,
+                    split_size=split_size,
+                    num_heads=num_heads // 2,
+                    dim_out=dim // 2,
+                    qk_scale=qk_scale,
+                    attn_drop=attn_drop,
+                    proj_drop=drop,
+                    position_bias=True,
+                )
+                for i in range(self.branch_num)
+            ]
+        )
+
+        if (self.rg_idx % 2 == 0 and self.b_idx > 0 and (self.b_idx - 2) % 4 == 0) or (
+            self.rg_idx % 2 != 0 and self.b_idx % 4 == 0
+        ):
+            attn_mask = self.calculate_mask(
+                self.patches_resolution, self.patches_resolution
+            )
+            self.register_buffer("attn_mask_0", attn_mask[0])
+            self.register_buffer("attn_mask_1", attn_mask[1])
+        else:
+            attn_mask = None
+            self.register_buffer("attn_mask_0", None)
+            self.register_buffer("attn_mask_1", None)
+
+        self.dwconv = nn.Sequential(
+            nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim),
+            nn.BatchNorm2d(dim),
+            nn.GELU(),
+        )
+        self.channel_interaction = nn.Sequential(
+            nn.AdaptiveAvgPool2d(1),
+            nn.Conv2d(dim, dim // 8, kernel_size=1),
+            nn.BatchNorm2d(dim // 8),
+            nn.GELU(),
+            nn.Conv2d(dim // 8, dim, kernel_size=1),
+        )
+        self.spatial_interaction = nn.Sequential(
+            nn.Conv2d(dim, dim // 16, kernel_size=1),
+            nn.BatchNorm2d(dim // 16),
+            nn.GELU(),
+            nn.Conv2d(dim // 16, 1, kernel_size=1),
+        )
+
+    def calculate_mask(self, H, W):
+        # The implementation builds on Swin Transformer code https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py
+        # calculate attention mask for shift window
+        img_mask_0 = torch.zeros((1, H, W, 1))  # 1 H W 1 idx=0
+        img_mask_1 = torch.zeros((1, H, W, 1))  # 1 H W 1 idx=1
+        h_slices_0 = (
+            slice(0, -self.split_size[0]),
+            slice(-self.split_size[0], -self.shift_size[0]),
+            slice(-self.shift_size[0], None),
+        )
+        w_slices_0 = (
+            slice(0, -self.split_size[1]),
+            slice(-self.split_size[1], -self.shift_size[1]),
+            slice(-self.shift_size[1], None),
+        )
+
+        h_slices_1 = (
+            slice(0, -self.split_size[1]),
+            slice(-self.split_size[1], -self.shift_size[1]),
+            slice(-self.shift_size[1], None),
+        )
+        w_slices_1 = (
+            slice(0, -self.split_size[0]),
+            slice(-self.split_size[0], -self.shift_size[0]),
+            slice(-self.shift_size[0], None),
+        )
+        cnt = 0
+        for h in h_slices_0:
+            for w in w_slices_0:
+                img_mask_0[:, h, w, :] = cnt
+                cnt += 1
+        cnt = 0
+        for h in h_slices_1:
+            for w in w_slices_1:
+                img_mask_1[:, h, w, :] = cnt
+                cnt += 1
+
+        # calculate mask for window-0
+        img_mask_0 = img_mask_0.view(
+            1,
+            H // self.split_size[0],
+            self.split_size[0],
+            W // self.split_size[1],
+            self.split_size[1],
+            1,
+        )
+        img_mask_0 = (
+            img_mask_0.permute(0, 1, 3, 2, 4, 5)
+            .contiguous()
+            .view(-1, self.split_size[0], self.split_size[1], 1)
+        )  # nW, sw[0], sw[1], 1
+        mask_windows_0 = img_mask_0.view(-1, self.split_size[0] * self.split_size[1])
+        attn_mask_0 = mask_windows_0.unsqueeze(1) - mask_windows_0.unsqueeze(2)
+        attn_mask_0 = attn_mask_0.masked_fill(
+            attn_mask_0 != 0, float(-100.0)
+        ).masked_fill(attn_mask_0 == 0, float(0.0))
+
+        # calculate mask for window-1
+        img_mask_1 = img_mask_1.view(
+            1,
+            H // self.split_size[1],
+            self.split_size[1],
+            W // self.split_size[0],
+            self.split_size[0],
+            1,
+        )
+        img_mask_1 = (
+            img_mask_1.permute(0, 1, 3, 2, 4, 5)
+            .contiguous()
+            .view(-1, self.split_size[1], self.split_size[0], 1)
+        )  # nW, sw[1], sw[0], 1
+        mask_windows_1 = img_mask_1.view(-1, self.split_size[1] * self.split_size[0])
+        attn_mask_1 = mask_windows_1.unsqueeze(1) - mask_windows_1.unsqueeze(2)
+        attn_mask_1 = attn_mask_1.masked_fill(
+            attn_mask_1 != 0, float(-100.0)
+        ).masked_fill(attn_mask_1 == 0, float(0.0))
+
+        return attn_mask_0, attn_mask_1
+
+    def forward(self, x, H, W):
+        """
+        Input: x: (B, H*W, C), H, W
+        Output: x: (B, H*W, C)
+        """
+        B, L, C = x.shape
+        assert L == H * W, "flatten img_tokens has wrong size"
+
+        qkv = self.qkv(x).reshape(B, -1, 3, C).permute(2, 0, 1, 3)  # 3, B, HW, C
+        # V without partition
+        v = qkv[2].transpose(-2, -1).contiguous().view(B, C, H, W)
+
+        # image padding
+        max_split_size = max(self.split_size[0], self.split_size[1])
+        pad_l = pad_t = 0
+        pad_r = (max_split_size - W % max_split_size) % max_split_size
+        pad_b = (max_split_size - H % max_split_size) % max_split_size
+
+        qkv = qkv.reshape(3 * B, H, W, C).permute(0, 3, 1, 2)  # 3B C H W
+        qkv = (
+            F.pad(qkv, (pad_l, pad_r, pad_t, pad_b))
+            .reshape(3, B, C, -1)
+            .transpose(-2, -1)
+        )  # l r t b
+        _H = pad_b + H
+        _W = pad_r + W
+        _L = _H * _W
+
+        # window-0 and window-1 on split channels [C/2, C/2]; for square windows (e.g., 8x8), window-0 and window-1 can be merged
+        # shift in block: (0, 4, 8, ...), (2, 6, 10, ...), (0, 4, 8, ...), (2, 6, 10, ...), ...
+        if (self.rg_idx % 2 == 0 and self.b_idx > 0 and (self.b_idx - 2) % 4 == 0) or (
+            self.rg_idx % 2 != 0 and self.b_idx % 4 == 0
+        ):
+            qkv = qkv.view(3, B, _H, _W, C)
+            qkv_0 = torch.roll(
+                qkv[:, :, :, :, : C // 2],
+                shifts=(-self.shift_size[0], -self.shift_size[1]),
+                dims=(2, 3),
+            )
+            qkv_0 = qkv_0.view(3, B, _L, C // 2)
+            qkv_1 = torch.roll(
+                qkv[:, :, :, :, C // 2 :],
+                shifts=(-self.shift_size[1], -self.shift_size[0]),
+                dims=(2, 3),
+            )
+            qkv_1 = qkv_1.view(3, B, _L, C // 2)
+
+            if self.patches_resolution != _H or self.patches_resolution != _W:
+                mask_tmp = self.calculate_mask(_H, _W)
+                x1_shift = self.attns[0](qkv_0, _H, _W, mask=mask_tmp[0].to(x.device))
+                x2_shift = self.attns[1](qkv_1, _H, _W, mask=mask_tmp[1].to(x.device))
+            else:
+                x1_shift = self.attns[0](qkv_0, _H, _W, mask=self.attn_mask_0)
+                x2_shift = self.attns[1](qkv_1, _H, _W, mask=self.attn_mask_1)
+
+            x1 = torch.roll(
+                x1_shift, shifts=(self.shift_size[0], self.shift_size[1]), dims=(1, 2)
+            )
+            x2 = torch.roll(
+                x2_shift, shifts=(self.shift_size[1], self.shift_size[0]), dims=(1, 2)
+            )
+            x1 = x1[:, :H, :W, :].reshape(B, L, C // 2)
+            x2 = x2[:, :H, :W, :].reshape(B, L, C // 2)
+            # attention output
+            attened_x = torch.cat([x1, x2], dim=2)
+
+        else:
+            x1 = self.attns[0](qkv[:, :, :, : C // 2], _H, _W)[:, :H, :W, :].reshape(
+                B, L, C // 2
+            )
+            x2 = self.attns[1](qkv[:, :, :, C // 2 :], _H, _W)[:, :H, :W, :].reshape(
+                B, L, C // 2
+            )
+            # attention output
+            attened_x = torch.cat([x1, x2], dim=2)
+
+        # convolution output
+        conv_x = self.dwconv(v)
+
+        # Adaptive Interaction Module (AIM)
+        # C-Map (before sigmoid)
+        channel_map = (
+            self.channel_interaction(conv_x)
+            .permute(0, 2, 3, 1)
+            .contiguous()
+            .view(B, 1, C)
+        )
+        # S-Map (before sigmoid)
+        attention_reshape = attened_x.transpose(-2, -1).contiguous().view(B, C, H, W)
+        spatial_map = self.spatial_interaction(attention_reshape)
+
+        # C-I
+        attened_x = attened_x * torch.sigmoid(channel_map)
+        # S-I
+        conv_x = torch.sigmoid(spatial_map) * conv_x
+        conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(B, L, C)
+
+        x = attened_x + conv_x
+
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+
+class Adaptive_Channel_Attention(nn.Module):
+    # The implementation builds on XCiT code https://github.com/facebookresearch/xcit
+    """Adaptive Channel Self-Attention
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads. Default: 6
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set.
+        attn_drop (float): Attention dropout rate. Default: 0.0
+        drop_path (float): Stochastic depth rate. Default: 0.0
+    """
+
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        qkv_bias=False,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1))
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.dwconv = nn.Sequential(
+            nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim),
+            nn.BatchNorm2d(dim),
+            nn.GELU(),
+        )
+        self.channel_interaction = nn.Sequential(
+            nn.AdaptiveAvgPool2d(1),
+            nn.Conv2d(dim, dim // 8, kernel_size=1),
+            nn.BatchNorm2d(dim // 8),
+            nn.GELU(),
+            nn.Conv2d(dim // 8, dim, kernel_size=1),
+        )
+        self.spatial_interaction = nn.Sequential(
+            nn.Conv2d(dim, dim // 16, kernel_size=1),
+            nn.BatchNorm2d(dim // 16),
+            nn.GELU(),
+            nn.Conv2d(dim // 16, 1, kernel_size=1),
+        )
+
+    def forward(self, x, H, W):
+        """
+        Input: x: (B, H*W, C), H, W
+        Output: x: (B, H*W, C)
+        """
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
+        qkv = qkv.permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        q = q.transpose(-2, -1)
+        k = k.transpose(-2, -1)
+        v = v.transpose(-2, -1)
+
+        v_ = v.reshape(B, C, N).contiguous().view(B, C, H, W)
+
+        q = torch.nn.functional.normalize(q, dim=-1)
+        k = torch.nn.functional.normalize(k, dim=-1)
+
+        attn = (q @ k.transpose(-2, -1)) * self.temperature
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        # attention output
+        attened_x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C)
+
+        # convolution output
+        conv_x = self.dwconv(v_)
+
+        # Adaptive Interaction Module (AIM)
+        # C-Map (before sigmoid)
+        attention_reshape = attened_x.transpose(-2, -1).contiguous().view(B, C, H, W)
+        channel_map = self.channel_interaction(attention_reshape)
+        # S-Map (before sigmoid)
+        spatial_map = (
+            self.spatial_interaction(conv_x)
+            .permute(0, 2, 3, 1)
+            .contiguous()
+            .view(B, N, 1)
+        )
+
+        # S-I
+        attened_x = attened_x * torch.sigmoid(spatial_map)
+        # C-I
+        conv_x = conv_x * torch.sigmoid(channel_map)
+        conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(B, N, C)
+
+        x = attened_x + conv_x
+
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+
+class DATB(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        reso=64,
+        split_size=[2, 4],
+        shift_size=[1, 2],
+        expansion_factor=4.0,
+        qkv_bias=False,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+        rg_idx=0,
+        b_idx=0,
+    ):
+        super().__init__()
+
+        self.norm1 = norm_layer(dim)
+
+        if b_idx % 2 == 0:
+            # DSTB
+            self.attn = Adaptive_Spatial_Attention(
+                dim,
+                num_heads=num_heads,
+                reso=reso,
+                split_size=split_size,
+                shift_size=shift_size,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop,
+                attn_drop=attn_drop,
+                rg_idx=rg_idx,
+                b_idx=b_idx,
+            )
+        else:
+            # DCTB
+            self.attn = Adaptive_Channel_Attention(
+                dim,
+                num_heads=num_heads,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                attn_drop=attn_drop,
+                proj_drop=drop,
+            )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        ffn_hidden_dim = int(dim * expansion_factor)
+        self.ffn = SGFN(
+            in_features=dim,
+            hidden_features=ffn_hidden_dim,
+            out_features=dim,
+            act_layer=act_layer,
+        )
+        self.norm2 = norm_layer(dim)
+
+    def forward(self, x, x_size):
+        """
+        Input: x: (B, H*W, C), x_size: (H, W)
+        Output: x: (B, H*W, C)
+        """
+        H, W = x_size
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.ffn(self.norm2(x), H, W))
+
+        return x
+
+
+class ResidualGroup(nn.Module):
+    """ResidualGroup
+    Args:
+        dim (int): Number of input channels.
+        reso (int): Input resolution.
+        num_heads (int): Number of attention heads.
+        split_size (tuple(int)): Height and Width of spatial window.
+        expansion_factor (float): Ratio of ffn hidden dim to embedding dim.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop (float): Dropout rate. Default: 0
+        attn_drop(float): Attention dropout rate. Default: 0
+        drop_paths (float | None): Stochastic depth rate.
+        act_layer (nn.Module): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm
+        depth (int): Number of dual aggregation Transformer blocks in residual group.
+        use_chk (bool): Whether to use checkpointing to save memory.
+        resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
+    """
+
+    def __init__(
+        self,
+        dim,
+        reso,
+        num_heads,
+        split_size=[2, 4],
+        expansion_factor=4.0,
+        qkv_bias=False,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_paths=None,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+        depth=2,
+        use_chk=False,
+        resi_connection="1conv",
+        rg_idx=0,
+    ):
+        super().__init__()
+        self.use_chk = use_chk
+        self.reso = reso
+
+        self.blocks = nn.ModuleList(
+            [
+                DATB(
+                    dim=dim,
+                    num_heads=num_heads,
+                    reso=reso,
+                    split_size=split_size,
+                    shift_size=[split_size[0] // 2, split_size[1] // 2],
+                    expansion_factor=expansion_factor,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop,
+                    attn_drop=attn_drop,
+                    drop_path=drop_paths[i],
+                    act_layer=act_layer,
+                    norm_layer=norm_layer,
+                    rg_idx=rg_idx,
+                    b_idx=i,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        if resi_connection == "1conv":
+            self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
+        elif resi_connection == "3conv":
+            self.conv = nn.Sequential(
+                nn.Conv2d(dim, dim // 4, 3, 1, 1),
+                nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
+                nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(dim // 4, dim, 3, 1, 1),
+            )
+
+    def forward(self, x, x_size):
+        """
+        Input: x: (B, H*W, C), x_size: (H, W)
+        Output: x: (B, H*W, C)
+        """
+        H, W = x_size
+        res = x
+        for blk in self.blocks:
+            if self.use_chk:
+                x = checkpoint.checkpoint(blk, x, x_size)
+            else:
+                x = blk(x, x_size)
+        x = rearrange(x, "b (h w) c -> b c h w", h=H, w=W)
+        x = self.conv(x)
+        x = rearrange(x, "b c h w -> b (h w) c")
+        x = res + x
+
+        return x
+
+
+class Upsample(nn.Sequential):
+    """Upsample module.
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(
+                f"scale {scale} is not supported. " "Supported scales: 2^n and 3."
+            )
+        super(Upsample, self).__init__(*m)
+
+
+class UpsampleOneStep(nn.Sequential):
+    """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
+       Used in lightweight SR to save parameters.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+
+    """
+
+    def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
+        self.num_feat = num_feat
+        self.input_resolution = input_resolution
+        m = []
+        m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1))
+        m.append(nn.PixelShuffle(scale))
+        super(UpsampleOneStep, self).__init__(*m)
+
+    def flops(self):
+        h, w = self.input_resolution
+        flops = h * w * self.num_feat * 3 * 9
+        return flops
+
+
+class DAT(nn.Module):
+    """Dual Aggregation Transformer
+    Args:
+        img_size (int): Input image size. Default: 64
+        in_chans (int): Number of input image channels. Default: 3
+        embed_dim (int): Patch embedding dimension. Default: 180
+        depths (tuple(int)): Depth of each residual group (number of DATB in each RG).
+        split_size (tuple(int)): Height and Width of spatial window.
+        num_heads (tuple(int)): Number of attention heads in different residual groups.
+        expansion_factor (float): Ratio of ffn hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        act_layer (nn.Module): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm
+        use_chk (bool): Whether to use checkpointing to save memory.
+        upscale: Upscale factor. 2/3/4 for image SR
+        img_range: Image range. 1. or 255.
+        resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
+    """
+
+    def __init__(self, state_dict):
+        super().__init__()
+
+        # defaults
+        img_size = 64
+        in_chans = 3
+        embed_dim = 180
+        split_size = [2, 4]
+        depth = [2, 2, 2, 2]
+        num_heads = [2, 2, 2, 2]
+        expansion_factor = 4.0
+        qkv_bias = True
+        qk_scale = None
+        drop_rate = 0.0
+        attn_drop_rate = 0.0
+        drop_path_rate = 0.1
+        act_layer = nn.GELU
+        norm_layer = nn.LayerNorm
+        use_chk = False
+        upscale = 2
+        img_range = 1.0
+        resi_connection = "1conv"
+        upsampler = "pixelshuffle"
+
+        self.model_arch = "DAT"
+        self.sub_type = "SR"
+        self.state = state_dict
+
+        state_keys = state_dict.keys()
+        if "conv_before_upsample.0.weight" in state_keys:
+            if "conv_up1.weight" in state_keys:
+                upsampler = "nearest+conv"
+            else:
+                upsampler = "pixelshuffle"
+                supports_fp16 = False
+        elif "upsample.0.weight" in state_keys:
+            upsampler = "pixelshuffledirect"
+        else:
+            upsampler = ""
+
+        num_feat = (
+            state_dict.get("conv_before_upsample.0.weight", None).shape[1]
+            if state_dict.get("conv_before_upsample.weight", None)
+            else 64
+        )
+
+        num_in_ch = state_dict["conv_first.weight"].shape[1]
+        in_chans = num_in_ch
+        if "conv_last.weight" in state_keys:
+            num_out_ch = state_dict["conv_last.weight"].shape[0]
+        else:
+            num_out_ch = num_in_ch
+
+        upscale = 1
+        if upsampler == "nearest+conv":
+            upsample_keys = [
+                x for x in state_keys if "conv_up" in x and "bias" not in x
+            ]
+
+            for upsample_key in upsample_keys:
+                upscale *= 2
+        elif upsampler == "pixelshuffle":
+            upsample_keys = [
+                x
+                for x in state_keys
+                if "upsample" in x and "conv" not in x and "bias" not in x
+            ]
+            for upsample_key in upsample_keys:
+                shape = state_dict[upsample_key].shape[0]
+                upscale *= math.sqrt(shape // num_feat)
+            upscale = int(upscale)
+        elif upsampler == "pixelshuffledirect":
+            upscale = int(
+                math.sqrt(state_dict["upsample.0.bias"].shape[0] // num_out_ch)
+            )
+
+        max_layer_num = 0
+        max_block_num = 0
+        for key in state_keys:
+            result = re.match(r"layers.(\d*).blocks.(\d*).norm1.weight", key)
+            if result:
+                layer_num, block_num = result.groups()
+                max_layer_num = max(max_layer_num, int(layer_num))
+                max_block_num = max(max_block_num, int(block_num))
+
+        depth = [max_block_num + 1 for _ in range(max_layer_num + 1)]
+
+        if "layers.0.blocks.1.attn.temperature" in state_keys:
+            num_heads_num = state_dict["layers.0.blocks.1.attn.temperature"].shape[0]
+            num_heads = [num_heads_num for _ in range(max_layer_num + 1)]
+        else:
+            num_heads = depth
+
+        embed_dim = state_dict["conv_first.weight"].shape[0]
+        expansion_factor = float(
+            state_dict["layers.0.blocks.0.ffn.fc1.weight"].shape[0] / embed_dim
+        )
+
+        # TODO: could actually count the layers, but this should do
+        if "layers.0.conv.4.weight" in state_keys:
+            resi_connection = "3conv"
+        else:
+            resi_connection = "1conv"
+
+        if "layers.0.blocks.2.attn.attn_mask_0" in state_keys:
+            attn_mask_0_x, attn_mask_0_y, attn_mask_0_z = state_dict[
+                "layers.0.blocks.2.attn.attn_mask_0"
+            ].shape
+
+            img_size = int(math.sqrt(attn_mask_0_x * attn_mask_0_y))
+
+        if "layers.0.blocks.0.attn.attns.0.rpe_biases" in state_keys:
+            split_sizes = (
+                state_dict["layers.0.blocks.0.attn.attns.0.rpe_biases"][-1] + 1
+            )
+            split_size = [int(x) for x in split_sizes]
+
+        self.in_nc = num_in_ch
+        self.out_nc = num_out_ch
+        self.num_feat = num_feat
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.depth = depth
+        self.scale = upscale
+        self.upsampler = upsampler
+        self.img_size = img_size
+        self.img_range = img_range
+        self.expansion_factor = expansion_factor
+        self.resi_connection = resi_connection
+        self.split_size = split_size
+
+        self.supports_fp16 = False  # Too much weirdness to support this at the moment
+        self.supports_bfp16 = True
+        self.min_size_restriction = 16
+
+        num_in_ch = in_chans
+        num_out_ch = in_chans
+        num_feat = 64
+        self.img_range = img_range
+        if in_chans == 3:
+            rgb_mean = (0.4488, 0.4371, 0.4040)
+            self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
+        else:
+            self.mean = torch.zeros(1, 1, 1, 1)
+        self.upscale = upscale
+        self.upsampler = upsampler
+
+        # ------------------------- 1, Shallow Feature Extraction ------------------------- #
+        self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
+
+        # ------------------------- 2, Deep Feature Extraction ------------------------- #
+        self.num_layers = len(depth)
+        self.use_chk = use_chk
+        self.num_features = (
+            self.embed_dim
+        ) = embed_dim  # num_features for consistency with other models
+        heads = num_heads
+
+        self.before_RG = nn.Sequential(
+            Rearrange("b c h w -> b (h w) c"), nn.LayerNorm(embed_dim)
+        )
+
+        curr_dim = embed_dim
+        dpr = [
+            x.item() for x in torch.linspace(0, drop_path_rate, np.sum(depth))
+        ]  # stochastic depth decay rule
+
+        self.layers = nn.ModuleList()
+        for i in range(self.num_layers):
+            layer = ResidualGroup(
+                dim=embed_dim,
+                num_heads=heads[i],
+                reso=img_size,
+                split_size=split_size,
+                expansion_factor=expansion_factor,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_paths=dpr[sum(depth[:i]) : sum(depth[: i + 1])],
+                act_layer=act_layer,
+                norm_layer=norm_layer,
+                depth=depth[i],
+                use_chk=use_chk,
+                resi_connection=resi_connection,
+                rg_idx=i,
+            )
+            self.layers.append(layer)
+
+        self.norm = norm_layer(curr_dim)
+        # build the last conv layer in deep feature extraction
+        if resi_connection == "1conv":
+            self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
+        elif resi_connection == "3conv":
+            # to save parameters and memory
+            self.conv_after_body = nn.Sequential(
+                nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
+                nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
+                nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1),
+            )
+
+        # ------------------------- 3, Reconstruction ------------------------- #
+        if self.upsampler == "pixelshuffle":
+            # for classical SR
+            self.conv_before_upsample = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
+            )
+            self.upsample = Upsample(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+        elif self.upsampler == "pixelshuffledirect":
+            # for lightweight SR (to save parameters)
+            self.upsample = UpsampleOneStep(
+                upscale, embed_dim, num_out_ch, (img_size, img_size)
+            )
+
+        self.apply(self._init_weights)
+        self.load_state_dict(state_dict, strict=True)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(
+            m, (nn.LayerNorm, nn.BatchNorm2d, nn.GroupNorm, nn.InstanceNorm2d)
+        ):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def forward_features(self, x):
+        _, _, H, W = x.shape
+        x_size = [H, W]
+        x = self.before_RG(x)
+        for layer in self.layers:
+            x = layer(x, x_size)
+        x = self.norm(x)
+        x = rearrange(x, "b (h w) c -> b c h w", h=H, w=W)
+
+        return x
+
+    def forward(self, x):
+        """
+        Input: x: (B, C, H, W)
+        """
+        self.mean = self.mean.type_as(x)
+        x = (x - self.mean) * self.img_range
+
+        if self.upsampler == "pixelshuffle":
+            # for image SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.conv_last(self.upsample(x))
+        elif self.upsampler == "pixelshuffledirect":
+            # for lightweight SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.upsample(x)
+
+        x = x / self.img_range + self.mean
+        return x
diff --git a/comfy_extras/chainner_models/architecture/LICENSE-DAT b/comfy_extras/chainner_models/architecture/LICENSE-DAT
new file mode 100644
index 00000000..261eeb9e
--- /dev/null
+++ b/comfy_extras/chainner_models/architecture/LICENSE-DAT
@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
diff --git a/comfy_extras/chainner_models/architecture/LICENSE-SCUNet b/comfy_extras/chainner_models/architecture/LICENSE-SCUNet
new file mode 100644
index 00000000..ff75c988
--- /dev/null
+++ b/comfy_extras/chainner_models/architecture/LICENSE-SCUNet
@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2022 Kai Zhang (cskaizhang@gmail.com, https://cszn.github.io/). All rights reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
diff --git a/comfy_extras/chainner_models/architecture/LICENSE-mat b/comfy_extras/chainner_models/architecture/LICENSE-mat
deleted file mode 100644
index 593adf6c..00000000
--- a/comfy_extras/chainner_models/architecture/LICENSE-mat
+++ /dev/null
@@ -1,161 +0,0 @@
-## creative commons
-
-# Attribution-NonCommercial 4.0 International
-
-Creative Commons Corporation (“Creative Commons”) is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer-client or other relationship. Creative Commons makes its licenses and related information available on an “as-is” basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible.
-
-### Using Creative Commons Public Licenses
-
-Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses.
-
-* __Considerations for licensors:__ Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC-licensed material, or material used under an exception or limitation to copyright. [More considerations for licensors](http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensors).
-
-* __Considerations for the public:__ By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor’s permission is not necessary for any reason–for example, because of any applicable exception or limitation to copyright–then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. [More considerations for the public](http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensees).
-
-## Creative Commons Attribution-NonCommercial 4.0 International Public License
-
-By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution-NonCommercial 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions.
-
-### Section 1 – Definitions.
-
-a. __Adapted Material__ means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image.
-
-b. __Adapter's License__ means the license You apply to Your Copyright and Similar Rights in Your contributions to Adapted Material in accordance with the terms and conditions of this Public License.
-
-c. __Copyright and Similar Rights__ means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)-(2) are not Copyright and Similar Rights.
-
-d. __Effective Technological Measures__ means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements.
-
-e. __Exceptions and Limitations__ means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material.
-
-f. __Licensed Material__ means the artistic or literary work, database, or other material to which the Licensor applied this Public License.
-
-g. __Licensed Rights__ means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license.
-
-h. __Licensor__ means the individual(s) or entity(ies) granting rights under this Public License.
-
-i. __NonCommercial__ means not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material subject to Copyright and Similar Rights by digital file-sharing or similar means is NonCommercial provided there is no payment of monetary compensation in connection with the exchange.
-
-j. __Share__ means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them.
-
-k. __Sui Generis Database Rights__ means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world.
-
-l. __You__ means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning.
-
-### Section 2 – Scope.
-
-a. ___License grant.___
-
-   1. Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to:
-
-       A. reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and
-
-       B. produce, reproduce, and Share Adapted Material for NonCommercial purposes only.
-
-   2. __Exceptions and Limitations.__ For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions.
-
-   3. __Term.__ The term of this Public License is specified in Section 6(a).
-
-   4. __Media and formats; technical modifications allowed.__ The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a)(4) never produces Adapted Material.
-
-    5. __Downstream recipients.__
-
-        A. __Offer from the Licensor – Licensed Material.__ Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License.
-
-        B. __No downstream restrictions.__ You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material.
-
-    6. __No endorsement.__ Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i).
-
-b. ___Other rights.___
-
-   1. Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise.
-
-   2. Patent and trademark rights are not licensed under this Public License.
-
-   3. To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties, including when the Licensed Material is used other than for NonCommercial purposes.
-
-### Section 3 – License Conditions.
-
-Your exercise of the Licensed Rights is expressly made subject to the following conditions.
-
-a. ___Attribution.___
-
-   1. If You Share the Licensed Material (including in modified form), You must:
-
-       A. retain the following if it is supplied by the Licensor with the Licensed Material:
-
-         i. identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated);
-
-         ii. a copyright notice;
-
-         iii. a notice that refers to this Public License;
-
-         iv. a notice that refers to the disclaimer of warranties;
-
-         v. a URI or hyperlink to the Licensed Material to the extent reasonably practicable;
-
-       B. indicate if You modified the Licensed Material and retain an indication of any previous modifications; and
-
-       C. indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License.
-
-   2. You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information.
-
-   3. If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable.
-
-   4. If You Share Adapted Material You produce, the Adapter's License You apply must not prevent recipients of the Adapted Material from complying with this Public License.
-
-### Section 4 – Sui Generis Database Rights.
-
-Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material:
-
-a. for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database for NonCommercial purposes only;
-
-b. if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material; and
-
-c. You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database.
-
-For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights.
-
-### Section 5 – Disclaimer of Warranties and Limitation of Liability.
-
-a. __Unless otherwise separately undertaken by the Licensor, to the extent possible, the Licensor offers the Licensed Material as-is and as-available, and makes no representations or warranties of any kind concerning the Licensed Material, whether express, implied, statutory, or other. This includes, without limitation, warranties of title, merchantability, fitness for a particular purpose, non-infringement, absence of latent or other defects, accuracy, or the presence or absence of errors, whether or not known or discoverable. Where disclaimers of warranties are not allowed in full or in part, this disclaimer may not apply to You.__
-
-b. __To the extent possible, in no event will the Licensor be liable to You on any legal theory (including, without limitation, negligence) or otherwise for any direct, special, indirect, incidental, consequential, punitive, exemplary, or other losses, costs, expenses, or damages arising out of this Public License or use of the Licensed Material, even if the Licensor has been advised of the possibility of such losses, costs, expenses, or damages. Where a limitation of liability is not allowed in full or in part, this limitation may not apply to You.__
-
-c. The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability.
-
-### Section 6 – Term and Termination.
-
-a. This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically.
-
-b. Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates:
-
-   1. automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or
-
-   2. upon express reinstatement by the Licensor.
-
-   For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License.
-
-c. For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License.
-
-d. Sections 1, 5, 6, 7, and 8 survive termination of this Public License.
-
-### Section 7 – Other Terms and Conditions.
-
-a. The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed.
-
-b. Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License.
-
-### Section 8 – Interpretation.
-
-a. For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License.
-
-b. To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions.
-
-c. No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor.
-
-d. Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority.
-
-> Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the “Licensor.” Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at [creativecommons.org/policies](http://creativecommons.org/policies), Creative Commons does not authorize the use of the trademark “Creative Commons” or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses.
->
-> Creative Commons may be contacted at creativecommons.org
diff --git a/comfy_extras/chainner_models/architecture/MAT.py b/comfy_extras/chainner_models/architecture/MAT.py
deleted file mode 100644
index 8fe17026..00000000
--- a/comfy_extras/chainner_models/architecture/MAT.py
+++ /dev/null
@@ -1,1636 +0,0 @@
-# pylint: skip-file
-"""Original MAT project is copyright of fenglingwb: https://github.com/fenglinglwb/MAT
-Code used for this implementation of MAT is modified from lama-cleaner,
-copyright of Sanster: https://github.com/fenglinglwb/MAT"""
-
-import random
-
-import numpy as np
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import torch.utils.checkpoint as checkpoint
-
-from .mat.utils import (
-    Conv2dLayer,
-    FullyConnectedLayer,
-    activation_funcs,
-    bias_act,
-    conv2d_resample,
-    normalize_2nd_moment,
-    setup_filter,
-    to_2tuple,
-    upsample2d,
-)
-
-
-class ModulatedConv2d(nn.Module):
-    def __init__(
-        self,
-        in_channels,  # Number of input channels.
-        out_channels,  # Number of output channels.
-        kernel_size,  # Width and height of the convolution kernel.
-        style_dim,  # dimension of the style code
-        demodulate=True,  # perfrom demodulation
-        up=1,  # Integer upsampling factor.
-        down=1,  # Integer downsampling factor.
-        resample_filter=[
-            1,
-            3,
-            3,
-            1,
-        ],  # Low-pass filter to apply when resampling activations.
-        conv_clamp=None,  # Clamp the output to +-X, None = disable clamping.
-    ):
-        super().__init__()
-        self.demodulate = demodulate
-
-        self.weight = torch.nn.Parameter(
-            torch.randn([1, out_channels, in_channels, kernel_size, kernel_size])
-        )
-        self.out_channels = out_channels
-        self.kernel_size = kernel_size
-        self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size**2))
-        self.padding = self.kernel_size // 2
-        self.up = up
-        self.down = down
-        self.register_buffer("resample_filter", setup_filter(resample_filter))
-        self.conv_clamp = conv_clamp
-
-        self.affine = FullyConnectedLayer(style_dim, in_channels, bias_init=1)
-
-    def forward(self, x, style):
-        batch, in_channels, height, width = x.shape
-        style = self.affine(style).view(batch, 1, in_channels, 1, 1).to(x.device)
-        weight = self.weight.to(x.device) * self.weight_gain * style
-
-        if self.demodulate:
-            decoefs = (weight.pow(2).sum(dim=[2, 3, 4]) + 1e-8).rsqrt()
-            weight = weight * decoefs.view(batch, self.out_channels, 1, 1, 1)
-
-        weight = weight.view(
-            batch * self.out_channels, in_channels, self.kernel_size, self.kernel_size
-        )
-        x = x.view(1, batch * in_channels, height, width)
-        x = conv2d_resample(
-            x=x,
-            w=weight,
-            f=self.resample_filter,
-            up=self.up,
-            down=self.down,
-            padding=self.padding,
-            groups=batch,
-        )
-        out = x.view(batch, self.out_channels, *x.shape[2:])
-
-        return out
-
-
-class StyleConv(torch.nn.Module):
-    def __init__(
-        self,
-        in_channels,  # Number of input channels.
-        out_channels,  # Number of output channels.
-        style_dim,  # Intermediate latent (W) dimensionality.
-        resolution,  # Resolution of this layer.
-        kernel_size=3,  # Convolution kernel size.
-        up=1,  # Integer upsampling factor.
-        use_noise=False,  # Enable noise input?
-        activation="lrelu",  # Activation function: 'relu', 'lrelu', etc.
-        resample_filter=[
-            1,
-            3,
-            3,
-            1,
-        ],  # Low-pass filter to apply when resampling activations.
-        conv_clamp=None,  # Clamp the output of convolution layers to +-X, None = disable clamping.
-        demodulate=True,  # perform demodulation
-    ):
-        super().__init__()
-
-        self.conv = ModulatedConv2d(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=kernel_size,
-            style_dim=style_dim,
-            demodulate=demodulate,
-            up=up,
-            resample_filter=resample_filter,
-            conv_clamp=conv_clamp,
-        )
-
-        self.use_noise = use_noise
-        self.resolution = resolution
-        if use_noise:
-            self.register_buffer("noise_const", torch.randn([resolution, resolution]))
-            self.noise_strength = torch.nn.Parameter(torch.zeros([]))
-
-        self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
-        self.activation = activation
-        self.act_gain = activation_funcs[activation].def_gain
-        self.conv_clamp = conv_clamp
-
-    def forward(self, x, style, noise_mode="random", gain=1):
-        x = self.conv(x, style)
-
-        assert noise_mode in ["random", "const", "none"]
-
-        if self.use_noise:
-            if noise_mode == "random":
-                xh, xw = x.size()[-2:]
-                noise = (
-                    torch.randn([x.shape[0], 1, xh, xw], device=x.device)
-                    * self.noise_strength
-                )
-            if noise_mode == "const":
-                noise = self.noise_const * self.noise_strength
-            x = x + noise
-
-        act_gain = self.act_gain * gain
-        act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
-        out = bias_act(
-            x, self.bias, act=self.activation, gain=act_gain, clamp=act_clamp
-        )
-
-        return out
-
-
-class ToRGB(torch.nn.Module):
-    def __init__(
-        self,
-        in_channels,
-        out_channels,
-        style_dim,
-        kernel_size=1,
-        resample_filter=[1, 3, 3, 1],
-        conv_clamp=None,
-        demodulate=False,
-    ):
-        super().__init__()
-
-        self.conv = ModulatedConv2d(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=kernel_size,
-            style_dim=style_dim,
-            demodulate=demodulate,
-            resample_filter=resample_filter,
-            conv_clamp=conv_clamp,
-        )
-        self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
-        self.register_buffer("resample_filter", setup_filter(resample_filter))
-        self.conv_clamp = conv_clamp
-
-    def forward(self, x, style, skip=None):
-        x = self.conv(x, style)
-        out = bias_act(x, self.bias, clamp=self.conv_clamp)
-
-        if skip is not None:
-            if skip.shape != out.shape:
-                skip = upsample2d(skip, self.resample_filter)
-            out = out + skip
-
-        return out
-
-
-def get_style_code(a, b):
-    return torch.cat([a, b.to(a.device)], dim=1)
-
-
-class DecBlockFirst(nn.Module):
-    def __init__(
-        self,
-        in_channels,
-        out_channels,
-        activation,
-        style_dim,
-        use_noise,
-        demodulate,
-        img_channels,
-    ):
-        super().__init__()
-        self.fc = FullyConnectedLayer(
-            in_features=in_channels * 2,
-            out_features=in_channels * 4**2,
-            activation=activation,
-        )
-        self.conv = StyleConv(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            style_dim=style_dim,
-            resolution=4,
-            kernel_size=3,
-            use_noise=use_noise,
-            activation=activation,
-            demodulate=demodulate,
-        )
-        self.toRGB = ToRGB(
-            in_channels=out_channels,
-            out_channels=img_channels,
-            style_dim=style_dim,
-            kernel_size=1,
-            demodulate=False,
-        )
-
-    def forward(self, x, ws, gs, E_features, noise_mode="random"):
-        x = self.fc(x).view(x.shape[0], -1, 4, 4)
-        x = x + E_features[2]
-        style = get_style_code(ws[:, 0], gs)
-        x = self.conv(x, style, noise_mode=noise_mode)
-        style = get_style_code(ws[:, 1], gs)
-        img = self.toRGB(x, style, skip=None)
-
-        return x, img
-
-
-class MappingNet(torch.nn.Module):
-    def __init__(
-        self,
-        z_dim,  # Input latent (Z) dimensionality, 0 = no latent.
-        c_dim,  # Conditioning label (C) dimensionality, 0 = no label.
-        w_dim,  # Intermediate latent (W) dimensionality.
-        num_ws,  # Number of intermediate latents to output, None = do not broadcast.
-        num_layers=8,  # Number of mapping layers.
-        embed_features=None,  # Label embedding dimensionality, None = same as w_dim.
-        layer_features=None,  # Number of intermediate features in the mapping layers, None = same as w_dim.
-        activation="lrelu",  # Activation function: 'relu', 'lrelu', etc.
-        lr_multiplier=0.01,  # Learning rate multiplier for the mapping layers.
-        w_avg_beta=0.995,  # Decay for tracking the moving average of W during training, None = do not track.
-    ):
-        super().__init__()
-        self.z_dim = z_dim
-        self.c_dim = c_dim
-        self.w_dim = w_dim
-        self.num_ws = num_ws
-        self.num_layers = num_layers
-        self.w_avg_beta = w_avg_beta
-
-        if embed_features is None:
-            embed_features = w_dim
-        if c_dim == 0:
-            embed_features = 0
-        if layer_features is None:
-            layer_features = w_dim
-        features_list = (
-            [z_dim + embed_features] + [layer_features] * (num_layers - 1) + [w_dim]
-        )
-
-        if c_dim > 0:
-            self.embed = FullyConnectedLayer(c_dim, embed_features)
-        for idx in range(num_layers):
-            in_features = features_list[idx]
-            out_features = features_list[idx + 1]
-            layer = FullyConnectedLayer(
-                in_features,
-                out_features,
-                activation=activation,
-                lr_multiplier=lr_multiplier,
-            )
-            setattr(self, f"fc{idx}", layer)
-
-        if num_ws is not None and w_avg_beta is not None:
-            self.register_buffer("w_avg", torch.zeros([w_dim]))
-
-    def forward(
-        self, z, c, truncation_psi=1, truncation_cutoff=None, skip_w_avg_update=False
-    ):
-        # Embed, normalize, and concat inputs.
-        x = None
-        with torch.autograd.profiler.record_function("input"):
-            if self.z_dim > 0:
-                x = normalize_2nd_moment(z.to(torch.float32))
-            if self.c_dim > 0:
-                y = normalize_2nd_moment(self.embed(c.to(torch.float32)))
-                x = torch.cat([x, y], dim=1) if x is not None else y
-
-        # Main layers.
-        for idx in range(self.num_layers):
-            layer = getattr(self, f"fc{idx}")
-            x = layer(x)
-
-        # Update moving average of W.
-        if self.w_avg_beta is not None and self.training and not skip_w_avg_update:
-            with torch.autograd.profiler.record_function("update_w_avg"):
-                self.w_avg.copy_(
-                    x.detach().mean(dim=0).lerp(self.w_avg, self.w_avg_beta)
-                )
-
-        # Broadcast.
-        if self.num_ws is not None:
-            with torch.autograd.profiler.record_function("broadcast"):
-                x = x.unsqueeze(1).repeat([1, self.num_ws, 1])
-
-        # Apply truncation.
-        if truncation_psi != 1:
-            with torch.autograd.profiler.record_function("truncate"):
-                assert self.w_avg_beta is not None
-                if self.num_ws is None or truncation_cutoff is None:
-                    x = self.w_avg.lerp(x, truncation_psi)
-                else:
-                    x[:, :truncation_cutoff] = self.w_avg.lerp(
-                        x[:, :truncation_cutoff], truncation_psi
-                    )
-
-        return x
-
-
-class DisFromRGB(nn.Module):
-    def __init__(
-        self, in_channels, out_channels, activation
-    ):  # res = 2, ..., resolution_log2
-        super().__init__()
-        self.conv = Conv2dLayer(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=1,
-            activation=activation,
-        )
-
-    def forward(self, x):
-        return self.conv(x)
-
-
-class DisBlock(nn.Module):
-    def __init__(
-        self, in_channels, out_channels, activation
-    ):  # res = 2, ..., resolution_log2
-        super().__init__()
-        self.conv0 = Conv2dLayer(
-            in_channels=in_channels,
-            out_channels=in_channels,
-            kernel_size=3,
-            activation=activation,
-        )
-        self.conv1 = Conv2dLayer(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=3,
-            down=2,
-            activation=activation,
-        )
-        self.skip = Conv2dLayer(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=1,
-            down=2,
-            bias=False,
-        )
-
-    def forward(self, x):
-        skip = self.skip(x, gain=np.sqrt(0.5))
-        x = self.conv0(x)
-        x = self.conv1(x, gain=np.sqrt(0.5))
-        out = skip + x
-
-        return out
-
-
-def nf(stage, channel_base=32768, channel_decay=1.0, channel_max=512):
-    NF = {512: 64, 256: 128, 128: 256, 64: 512, 32: 512, 16: 512, 8: 512, 4: 512}
-    return NF[2**stage]
-
-
-class Mlp(nn.Module):
-    def __init__(
-        self,
-        in_features,
-        hidden_features=None,
-        out_features=None,
-        act_layer=nn.GELU,
-        drop=0.0,
-    ):
-        super().__init__()
-        out_features = out_features or in_features
-        hidden_features = hidden_features or in_features
-        self.fc1 = FullyConnectedLayer(
-            in_features=in_features, out_features=hidden_features, activation="lrelu"
-        )
-        self.fc2 = FullyConnectedLayer(
-            in_features=hidden_features, out_features=out_features
-        )
-
-    def forward(self, x):
-        x = self.fc1(x)
-        x = self.fc2(x)
-        return x
-
-
-def window_partition(x, window_size):
-    """
-    Args:
-        x: (B, H, W, C)
-        window_size (int): window size
-    Returns:
-        windows: (num_windows*B, window_size, window_size, C)
-    """
-    B, H, W, C = x.shape
-    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
-    windows = (
-        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
-    )
-    return windows
-
-
-def window_reverse(windows, window_size: int, H: int, W: int):
-    """
-    Args:
-        windows: (num_windows*B, window_size, window_size, C)
-        window_size (int): Window size
-        H (int): Height of image
-        W (int): Width of image
-    Returns:
-        x: (B, H, W, C)
-    """
-    B = int(windows.shape[0] / (H * W / window_size / window_size))
-    # B = windows.shape[0] / (H * W / window_size / window_size)
-    x = windows.view(
-        B, H // window_size, W // window_size, window_size, window_size, -1
-    )
-    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
-    return x
-
-
-class Conv2dLayerPartial(nn.Module):
-    def __init__(
-        self,
-        in_channels,  # Number of input channels.
-        out_channels,  # Number of output channels.
-        kernel_size,  # Width and height of the convolution kernel.
-        bias=True,  # Apply additive bias before the activation function?
-        activation="linear",  # Activation function: 'relu', 'lrelu', etc.
-        up=1,  # Integer upsampling factor.
-        down=1,  # Integer downsampling factor.
-        resample_filter=[
-            1,
-            3,
-            3,
-            1,
-        ],  # Low-pass filter to apply when resampling activations.
-        conv_clamp=None,  # Clamp the output to +-X, None = disable clamping.
-        trainable=True,  # Update the weights of this layer during training?
-    ):
-        super().__init__()
-        self.conv = Conv2dLayer(
-            in_channels,
-            out_channels,
-            kernel_size,
-            bias,
-            activation,
-            up,
-            down,
-            resample_filter,
-            conv_clamp,
-            trainable,
-        )
-
-        self.weight_maskUpdater = torch.ones(1, 1, kernel_size, kernel_size)
-        self.slide_winsize = kernel_size**2
-        self.stride = down
-        self.padding = kernel_size // 2 if kernel_size % 2 == 1 else 0
-
-    def forward(self, x, mask=None):
-        if mask is not None:
-            with torch.no_grad():
-                if self.weight_maskUpdater.type() != x.type():
-                    self.weight_maskUpdater = self.weight_maskUpdater.to(x)
-                update_mask = F.conv2d(
-                    mask,
-                    self.weight_maskUpdater,
-                    bias=None,
-                    stride=self.stride,
-                    padding=self.padding,
-                )
-                mask_ratio = self.slide_winsize / (update_mask + 1e-8)
-                update_mask = torch.clamp(update_mask, 0, 1)  # 0 or 1
-                mask_ratio = torch.mul(mask_ratio, update_mask)
-            x = self.conv(x)
-            x = torch.mul(x, mask_ratio)
-            return x, update_mask
-        else:
-            x = self.conv(x)
-            return x, None
-
-
-class WindowAttention(nn.Module):
-    r"""Window based multi-head self attention (W-MSA) module with relative position bias.
-    It supports both of shifted and non-shifted window.
-    Args:
-        dim (int): Number of input channels.
-        window_size (tuple[int]): The height and width of the window.
-        num_heads (int): Number of attention heads.
-        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
-        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
-        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
-        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
-    """
-
-    def __init__(
-        self,
-        dim,
-        window_size,
-        num_heads,
-        down_ratio=1,
-        qkv_bias=True,
-        qk_scale=None,
-        attn_drop=0.0,
-        proj_drop=0.0,
-    ):
-        super().__init__()
-        self.dim = dim
-        self.window_size = window_size  # Wh, Ww
-        self.num_heads = num_heads
-        head_dim = dim // num_heads
-        self.scale = qk_scale or head_dim**-0.5
-
-        self.q = FullyConnectedLayer(in_features=dim, out_features=dim)
-        self.k = FullyConnectedLayer(in_features=dim, out_features=dim)
-        self.v = FullyConnectedLayer(in_features=dim, out_features=dim)
-        self.proj = FullyConnectedLayer(in_features=dim, out_features=dim)
-
-        self.softmax = nn.Softmax(dim=-1)
-
-    def forward(self, x, mask_windows=None, mask=None):
-        """
-        Args:
-            x: input features with shape of (num_windows*B, N, C)
-            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
-        """
-        B_, N, C = x.shape
-        norm_x = F.normalize(x, p=2.0, dim=-1)
-        q = (
-            self.q(norm_x)
-            .reshape(B_, N, self.num_heads, C // self.num_heads)
-            .permute(0, 2, 1, 3)
-        )
-        k = (
-            self.k(norm_x)
-            .view(B_, -1, self.num_heads, C // self.num_heads)
-            .permute(0, 2, 3, 1)
-        )
-        v = (
-            self.v(x)
-            .view(B_, -1, self.num_heads, C // self.num_heads)
-            .permute(0, 2, 1, 3)
-        )
-
-        attn = (q @ k) * self.scale
-
-        if mask is not None:
-            nW = mask.shape[0]
-            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
-                1
-            ).unsqueeze(0)
-            attn = attn.view(-1, self.num_heads, N, N)
-
-        if mask_windows is not None:
-            attn_mask_windows = mask_windows.squeeze(-1).unsqueeze(1).unsqueeze(1)
-            attn = attn + attn_mask_windows.masked_fill(
-                attn_mask_windows == 0, float(-100.0)
-            ).masked_fill(attn_mask_windows == 1, float(0.0))
-            with torch.no_grad():
-                mask_windows = torch.clamp(
-                    torch.sum(mask_windows, dim=1, keepdim=True), 0, 1
-                ).repeat(1, N, 1)
-
-        attn = self.softmax(attn)
-
-        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
-        x = self.proj(x)
-        return x, mask_windows
-
-
-class SwinTransformerBlock(nn.Module):
-    r"""Swin Transformer Block.
-    Args:
-        dim (int): Number of input channels.
-        input_resolution (tuple[int]): Input resulotion.
-        num_heads (int): Number of attention heads.
-        window_size (int): Window size.
-        shift_size (int): Shift size for SW-MSA.
-        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
-        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
-        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
-        drop (float, optional): Dropout rate. Default: 0.0
-        attn_drop (float, optional): Attention dropout rate. Default: 0.0
-        drop_path (float, optional): Stochastic depth rate. Default: 0.0
-        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
-        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
-    """
-
-    def __init__(
-        self,
-        dim,
-        input_resolution,
-        num_heads,
-        down_ratio=1,
-        window_size=7,
-        shift_size=0,
-        mlp_ratio=4.0,
-        qkv_bias=True,
-        qk_scale=None,
-        drop=0.0,
-        attn_drop=0.0,
-        drop_path=0.0,
-        act_layer=nn.GELU,
-        norm_layer=nn.LayerNorm,
-    ):
-        super().__init__()
-        self.dim = dim
-        self.input_resolution = input_resolution
-        self.num_heads = num_heads
-        self.window_size = window_size
-        self.shift_size = shift_size
-        self.mlp_ratio = mlp_ratio
-        if min(self.input_resolution) <= self.window_size:
-            # if window size is larger than input resolution, we don't partition windows
-            self.shift_size = 0
-            self.window_size = min(self.input_resolution)
-        assert (
-            0 <= self.shift_size < self.window_size
-        ), "shift_size must in 0-window_size"
-
-        if self.shift_size > 0:
-            down_ratio = 1
-        self.attn = WindowAttention(
-            dim,
-            window_size=to_2tuple(self.window_size),
-            num_heads=num_heads,
-            down_ratio=down_ratio,
-            qkv_bias=qkv_bias,
-            qk_scale=qk_scale,
-            attn_drop=attn_drop,
-            proj_drop=drop,
-        )
-
-        self.fuse = FullyConnectedLayer(
-            in_features=dim * 2, out_features=dim, activation="lrelu"
-        )
-
-        mlp_hidden_dim = int(dim * mlp_ratio)
-        self.mlp = Mlp(
-            in_features=dim,
-            hidden_features=mlp_hidden_dim,
-            act_layer=act_layer,
-            drop=drop,
-        )
-
-        if self.shift_size > 0:
-            attn_mask = self.calculate_mask(self.input_resolution)
-        else:
-            attn_mask = None
-
-        self.register_buffer("attn_mask", attn_mask)
-
-    def calculate_mask(self, x_size):
-        # calculate attention mask for SW-MSA
-        H, W = x_size
-        img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
-        h_slices = (
-            slice(0, -self.window_size),
-            slice(-self.window_size, -self.shift_size),
-            slice(-self.shift_size, None),
-        )
-        w_slices = (
-            slice(0, -self.window_size),
-            slice(-self.window_size, -self.shift_size),
-            slice(-self.shift_size, None),
-        )
-        cnt = 0
-        for h in h_slices:
-            for w in w_slices:
-                img_mask[:, h, w, :] = cnt
-                cnt += 1
-
-        mask_windows = window_partition(
-            img_mask, self.window_size
-        )  # nW, window_size, window_size, 1
-        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
-        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
-        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
-            attn_mask == 0, float(0.0)
-        )
-
-        return attn_mask
-
-    def forward(self, x, x_size, mask=None):
-        # H, W = self.input_resolution
-        H, W = x_size
-        B, _, C = x.shape
-        # assert L == H * W, "input feature has wrong size"
-
-        shortcut = x
-        x = x.view(B, H, W, C)
-        if mask is not None:
-            mask = mask.view(B, H, W, 1)
-
-        # cyclic shift
-        if self.shift_size > 0:
-            shifted_x = torch.roll(
-                x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
-            )
-            if mask is not None:
-                shifted_mask = torch.roll(
-                    mask, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
-                )
-        else:
-            shifted_x = x
-            if mask is not None:
-                shifted_mask = mask
-
-        # partition windows
-        x_windows = window_partition(
-            shifted_x, self.window_size
-        )  # nW*B, window_size, window_size, C
-        x_windows = x_windows.view(
-            -1, self.window_size * self.window_size, C
-        )  # nW*B, window_size*window_size, C
-        if mask is not None:
-            mask_windows = window_partition(shifted_mask, self.window_size)
-            mask_windows = mask_windows.view(-1, self.window_size * self.window_size, 1)
-        else:
-            mask_windows = None
-
-        # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
-        if self.input_resolution == x_size:
-            attn_windows, mask_windows = self.attn(
-                x_windows, mask_windows, mask=self.attn_mask
-            )  # nW*B, window_size*window_size, C
-        else:
-            attn_windows, mask_windows = self.attn(
-                x_windows, mask_windows, mask=self.calculate_mask(x_size).to(x.device)
-            )  # nW*B, window_size*window_size, C
-
-        # merge windows
-        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
-        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
-        if mask is not None:
-            mask_windows = mask_windows.view(-1, self.window_size, self.window_size, 1)
-            shifted_mask = window_reverse(mask_windows, self.window_size, H, W)
-
-        # reverse cyclic shift
-        if self.shift_size > 0:
-            x = torch.roll(
-                shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
-            )
-            if mask is not None:
-                mask = torch.roll(
-                    shifted_mask, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
-                )
-        else:
-            x = shifted_x
-            if mask is not None:
-                mask = shifted_mask
-        x = x.view(B, H * W, C)
-        if mask is not None:
-            mask = mask.view(B, H * W, 1)
-
-        # FFN
-        x = self.fuse(torch.cat([shortcut, x], dim=-1))
-        x = self.mlp(x)
-
-        return x, mask
-
-
-class PatchMerging(nn.Module):
-    def __init__(self, in_channels, out_channels, down=2):
-        super().__init__()
-        self.conv = Conv2dLayerPartial(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=3,
-            activation="lrelu",
-            down=down,
-        )
-        self.down = down
-
-    def forward(self, x, x_size, mask=None):
-        x = token2feature(x, x_size)
-        if mask is not None:
-            mask = token2feature(mask, x_size)
-        x, mask = self.conv(x, mask)
-        if self.down != 1:
-            ratio = 1 / self.down
-            x_size = (int(x_size[0] * ratio), int(x_size[1] * ratio))
-        x = feature2token(x)
-        if mask is not None:
-            mask = feature2token(mask)
-        return x, x_size, mask
-
-
-class PatchUpsampling(nn.Module):
-    def __init__(self, in_channels, out_channels, up=2):
-        super().__init__()
-        self.conv = Conv2dLayerPartial(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=3,
-            activation="lrelu",
-            up=up,
-        )
-        self.up = up
-
-    def forward(self, x, x_size, mask=None):
-        x = token2feature(x, x_size)
-        if mask is not None:
-            mask = token2feature(mask, x_size)
-        x, mask = self.conv(x, mask)
-        if self.up != 1:
-            x_size = (int(x_size[0] * self.up), int(x_size[1] * self.up))
-        x = feature2token(x)
-        if mask is not None:
-            mask = feature2token(mask)
-        return x, x_size, mask
-
-
-class BasicLayer(nn.Module):
-    """A basic Swin Transformer layer for one stage.
-    Args:
-        dim (int): Number of input channels.
-        input_resolution (tuple[int]): Input resolution.
-        depth (int): Number of blocks.
-        num_heads (int): Number of attention heads.
-        window_size (int): Local window size.
-        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
-        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
-        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
-        drop (float, optional): Dropout rate. Default: 0.0
-        attn_drop (float, optional): Attention dropout rate. Default: 0.0
-        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
-        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
-        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
-        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
-    """
-
-    def __init__(
-        self,
-        dim,
-        input_resolution,
-        depth,
-        num_heads,
-        window_size,
-        down_ratio=1,
-        mlp_ratio=2.0,
-        qkv_bias=True,
-        qk_scale=None,
-        drop=0.0,
-        attn_drop=0.0,
-        drop_path=0.0,
-        norm_layer=nn.LayerNorm,
-        downsample=None,
-        use_checkpoint=False,
-    ):
-        super().__init__()
-        self.dim = dim
-        self.input_resolution = input_resolution
-        self.depth = depth
-        self.use_checkpoint = use_checkpoint
-
-        # patch merging layer
-        if downsample is not None:
-            # self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
-            self.downsample = downsample
-        else:
-            self.downsample = None
-
-        # build blocks
-        self.blocks = nn.ModuleList(
-            [
-                SwinTransformerBlock(
-                    dim=dim,
-                    input_resolution=input_resolution,
-                    num_heads=num_heads,
-                    down_ratio=down_ratio,
-                    window_size=window_size,
-                    shift_size=0 if (i % 2 == 0) else window_size // 2,
-                    mlp_ratio=mlp_ratio,
-                    qkv_bias=qkv_bias,
-                    qk_scale=qk_scale,
-                    drop=drop,
-                    attn_drop=attn_drop,
-                    drop_path=drop_path[i]
-                    if isinstance(drop_path, list)
-                    else drop_path,
-                    norm_layer=norm_layer,
-                )
-                for i in range(depth)
-            ]
-        )
-
-        self.conv = Conv2dLayerPartial(
-            in_channels=dim, out_channels=dim, kernel_size=3, activation="lrelu"
-        )
-
-    def forward(self, x, x_size, mask=None):
-        if self.downsample is not None:
-            x, x_size, mask = self.downsample(x, x_size, mask)
-        identity = x
-        for blk in self.blocks:
-            if self.use_checkpoint:
-                x, mask = checkpoint.checkpoint(blk, x, x_size, mask)
-            else:
-                x, mask = blk(x, x_size, mask)
-        if mask is not None:
-            mask = token2feature(mask, x_size)
-        x, mask = self.conv(token2feature(x, x_size), mask)
-        x = feature2token(x) + identity
-        if mask is not None:
-            mask = feature2token(mask)
-        return x, x_size, mask
-
-
-class ToToken(nn.Module):
-    def __init__(self, in_channels=3, dim=128, kernel_size=5, stride=1):
-        super().__init__()
-
-        self.proj = Conv2dLayerPartial(
-            in_channels=in_channels,
-            out_channels=dim,
-            kernel_size=kernel_size,
-            activation="lrelu",
-        )
-
-    def forward(self, x, mask):
-        x, mask = self.proj(x, mask)
-
-        return x, mask
-
-
-class EncFromRGB(nn.Module):
-    def __init__(
-        self, in_channels, out_channels, activation
-    ):  # res = 2, ..., resolution_log2
-        super().__init__()
-        self.conv0 = Conv2dLayer(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=1,
-            activation=activation,
-        )
-        self.conv1 = Conv2dLayer(
-            in_channels=out_channels,
-            out_channels=out_channels,
-            kernel_size=3,
-            activation=activation,
-        )
-
-    def forward(self, x):
-        x = self.conv0(x)
-        x = self.conv1(x)
-
-        return x
-
-
-class ConvBlockDown(nn.Module):
-    def __init__(
-        self, in_channels, out_channels, activation
-    ):  # res = 2, ..., resolution_log
-        super().__init__()
-
-        self.conv0 = Conv2dLayer(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            kernel_size=3,
-            activation=activation,
-            down=2,
-        )
-        self.conv1 = Conv2dLayer(
-            in_channels=out_channels,
-            out_channels=out_channels,
-            kernel_size=3,
-            activation=activation,
-        )
-
-    def forward(self, x):
-        x = self.conv0(x)
-        x = self.conv1(x)
-
-        return x
-
-
-def token2feature(x, x_size):
-    B, _, C = x.shape
-    h, w = x_size
-    x = x.permute(0, 2, 1).reshape(B, C, h, w)
-    return x
-
-
-def feature2token(x):
-    B, C, _, _ = x.shape
-    x = x.view(B, C, -1).transpose(1, 2)
-    return x
-
-
-class Encoder(nn.Module):
-    def __init__(
-        self,
-        res_log2,
-        img_channels,
-        activation,
-        patch_size=5,
-        channels=16,
-        drop_path_rate=0.1,
-    ):
-        super().__init__()
-
-        self.resolution = []
-
-        for i in range(res_log2, 3, -1):  # from input size to 16x16
-            res = 2**i
-            self.resolution.append(res)
-            if i == res_log2:
-                block = EncFromRGB(img_channels * 2 + 1, nf(i), activation)
-            else:
-                block = ConvBlockDown(nf(i + 1), nf(i), activation)
-            setattr(self, "EncConv_Block_%dx%d" % (res, res), block)
-
-    def forward(self, x):
-        out = {}
-        for res in self.resolution:
-            res_log2 = int(np.log2(res))
-            x = getattr(self, "EncConv_Block_%dx%d" % (res, res))(x)
-            out[res_log2] = x
-
-        return out
-
-
-class ToStyle(nn.Module):
-    def __init__(self, in_channels, out_channels, activation, drop_rate):
-        super().__init__()
-        self.conv = nn.Sequential(
-            Conv2dLayer(
-                in_channels=in_channels,
-                out_channels=in_channels,
-                kernel_size=3,
-                activation=activation,
-                down=2,
-            ),
-            Conv2dLayer(
-                in_channels=in_channels,
-                out_channels=in_channels,
-                kernel_size=3,
-                activation=activation,
-                down=2,
-            ),
-            Conv2dLayer(
-                in_channels=in_channels,
-                out_channels=in_channels,
-                kernel_size=3,
-                activation=activation,
-                down=2,
-            ),
-        )
-
-        self.pool = nn.AdaptiveAvgPool2d(1)
-        self.fc = FullyConnectedLayer(
-            in_features=in_channels, out_features=out_channels, activation=activation
-        )
-        # self.dropout = nn.Dropout(drop_rate)
-
-    def forward(self, x):
-        x = self.conv(x)
-        x = self.pool(x)
-        x = self.fc(x.flatten(start_dim=1))
-        # x = self.dropout(x)
-
-        return x
-
-
-class DecBlockFirstV2(nn.Module):
-    def __init__(
-        self,
-        res,
-        in_channels,
-        out_channels,
-        activation,
-        style_dim,
-        use_noise,
-        demodulate,
-        img_channels,
-    ):
-        super().__init__()
-        self.res = res
-
-        self.conv0 = Conv2dLayer(
-            in_channels=in_channels,
-            out_channels=in_channels,
-            kernel_size=3,
-            activation=activation,
-        )
-        self.conv1 = StyleConv(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            style_dim=style_dim,
-            resolution=2**res,
-            kernel_size=3,
-            use_noise=use_noise,
-            activation=activation,
-            demodulate=demodulate,
-        )
-        self.toRGB = ToRGB(
-            in_channels=out_channels,
-            out_channels=img_channels,
-            style_dim=style_dim,
-            kernel_size=1,
-            demodulate=False,
-        )
-
-    def forward(self, x, ws, gs, E_features, noise_mode="random"):
-        # x = self.fc(x).view(x.shape[0], -1, 4, 4)
-        x = self.conv0(x)
-        x = x + E_features[self.res]
-        style = get_style_code(ws[:, 0], gs)
-        x = self.conv1(x, style, noise_mode=noise_mode)
-        style = get_style_code(ws[:, 1], gs)
-        img = self.toRGB(x, style, skip=None)
-
-        return x, img
-
-
-class DecBlock(nn.Module):
-    def __init__(
-        self,
-        res,
-        in_channels,
-        out_channels,
-        activation,
-        style_dim,
-        use_noise,
-        demodulate,
-        img_channels,
-    ):  # res = 4, ..., resolution_log2
-        super().__init__()
-        self.res = res
-
-        self.conv0 = StyleConv(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            style_dim=style_dim,
-            resolution=2**res,
-            kernel_size=3,
-            up=2,
-            use_noise=use_noise,
-            activation=activation,
-            demodulate=demodulate,
-        )
-        self.conv1 = StyleConv(
-            in_channels=out_channels,
-            out_channels=out_channels,
-            style_dim=style_dim,
-            resolution=2**res,
-            kernel_size=3,
-            use_noise=use_noise,
-            activation=activation,
-            demodulate=demodulate,
-        )
-        self.toRGB = ToRGB(
-            in_channels=out_channels,
-            out_channels=img_channels,
-            style_dim=style_dim,
-            kernel_size=1,
-            demodulate=False,
-        )
-
-    def forward(self, x, img, ws, gs, E_features, noise_mode="random"):
-        style = get_style_code(ws[:, self.res * 2 - 9], gs)
-        x = self.conv0(x, style, noise_mode=noise_mode)
-        x = x + E_features[self.res]
-        style = get_style_code(ws[:, self.res * 2 - 8], gs)
-        x = self.conv1(x, style, noise_mode=noise_mode)
-        style = get_style_code(ws[:, self.res * 2 - 7], gs)
-        img = self.toRGB(x, style, skip=img)
-
-        return x, img
-
-
-class Decoder(nn.Module):
-    def __init__(
-        self, res_log2, activation, style_dim, use_noise, demodulate, img_channels
-    ):
-        super().__init__()
-        self.Dec_16x16 = DecBlockFirstV2(
-            4, nf(4), nf(4), activation, style_dim, use_noise, demodulate, img_channels
-        )
-        for res in range(5, res_log2 + 1):
-            setattr(
-                self,
-                "Dec_%dx%d" % (2**res, 2**res),
-                DecBlock(
-                    res,
-                    nf(res - 1),
-                    nf(res),
-                    activation,
-                    style_dim,
-                    use_noise,
-                    demodulate,
-                    img_channels,
-                ),
-            )
-        self.res_log2 = res_log2
-
-    def forward(self, x, ws, gs, E_features, noise_mode="random"):
-        x, img = self.Dec_16x16(x, ws, gs, E_features, noise_mode=noise_mode)
-        for res in range(5, self.res_log2 + 1):
-            block = getattr(self, "Dec_%dx%d" % (2**res, 2**res))
-            x, img = block(x, img, ws, gs, E_features, noise_mode=noise_mode)
-
-        return img
-
-
-class DecStyleBlock(nn.Module):
-    def __init__(
-        self,
-        res,
-        in_channels,
-        out_channels,
-        activation,
-        style_dim,
-        use_noise,
-        demodulate,
-        img_channels,
-    ):
-        super().__init__()
-        self.res = res
-
-        self.conv0 = StyleConv(
-            in_channels=in_channels,
-            out_channels=out_channels,
-            style_dim=style_dim,
-            resolution=2**res,
-            kernel_size=3,
-            up=2,
-            use_noise=use_noise,
-            activation=activation,
-            demodulate=demodulate,
-        )
-        self.conv1 = StyleConv(
-            in_channels=out_channels,
-            out_channels=out_channels,
-            style_dim=style_dim,
-            resolution=2**res,
-            kernel_size=3,
-            use_noise=use_noise,
-            activation=activation,
-            demodulate=demodulate,
-        )
-        self.toRGB = ToRGB(
-            in_channels=out_channels,
-            out_channels=img_channels,
-            style_dim=style_dim,
-            kernel_size=1,
-            demodulate=False,
-        )
-
-    def forward(self, x, img, style, skip, noise_mode="random"):
-        x = self.conv0(x, style, noise_mode=noise_mode)
-        x = x + skip
-        x = self.conv1(x, style, noise_mode=noise_mode)
-        img = self.toRGB(x, style, skip=img)
-
-        return x, img
-
-
-class FirstStage(nn.Module):
-    def __init__(
-        self,
-        img_channels,
-        img_resolution=256,
-        dim=180,
-        w_dim=512,
-        use_noise=False,
-        demodulate=True,
-        activation="lrelu",
-    ):
-        super().__init__()
-        res = 64
-
-        self.conv_first = Conv2dLayerPartial(
-            in_channels=img_channels + 1,
-            out_channels=dim,
-            kernel_size=3,
-            activation=activation,
-        )
-        self.enc_conv = nn.ModuleList()
-        down_time = int(np.log2(img_resolution // res))
-        # 根据图片尺寸构建 swim transformer 的层数
-        for i in range(down_time):  # from input size to 64
-            self.enc_conv.append(
-                Conv2dLayerPartial(
-                    in_channels=dim,
-                    out_channels=dim,
-                    kernel_size=3,
-                    down=2,
-                    activation=activation,
-                )
-            )
-
-        # from 64 -> 16 -> 64
-        depths = [2, 3, 4, 3, 2]
-        ratios = [1, 1 / 2, 1 / 2, 2, 2]
-        num_heads = 6
-        window_sizes = [8, 16, 16, 16, 8]
-        drop_path_rate = 0.1
-        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
-
-        self.tran = nn.ModuleList()
-        for i, depth in enumerate(depths):
-            res = int(res * ratios[i])
-            if ratios[i] < 1:
-                merge = PatchMerging(dim, dim, down=int(1 / ratios[i]))
-            elif ratios[i] > 1:
-                merge = PatchUpsampling(dim, dim, up=ratios[i])
-            else:
-                merge = None
-            self.tran.append(
-                BasicLayer(
-                    dim=dim,
-                    input_resolution=[res, res],
-                    depth=depth,
-                    num_heads=num_heads,
-                    window_size=window_sizes[i],
-                    drop_path=dpr[sum(depths[:i]) : sum(depths[: i + 1])],
-                    downsample=merge,
-                )
-            )
-
-        # global style
-        down_conv = []
-        for i in range(int(np.log2(16))):
-            down_conv.append(
-                Conv2dLayer(
-                    in_channels=dim,
-                    out_channels=dim,
-                    kernel_size=3,
-                    down=2,
-                    activation=activation,
-                )
-            )
-        down_conv.append(nn.AdaptiveAvgPool2d((1, 1)))
-        self.down_conv = nn.Sequential(*down_conv)
-        self.to_style = FullyConnectedLayer(
-            in_features=dim, out_features=dim * 2, activation=activation
-        )
-        self.ws_style = FullyConnectedLayer(
-            in_features=w_dim, out_features=dim, activation=activation
-        )
-        self.to_square = FullyConnectedLayer(
-            in_features=dim, out_features=16 * 16, activation=activation
-        )
-
-        style_dim = dim * 3
-        self.dec_conv = nn.ModuleList()
-        for i in range(down_time):  # from 64 to input size
-            res = res * 2
-            self.dec_conv.append(
-                DecStyleBlock(
-                    res,
-                    dim,
-                    dim,
-                    activation,
-                    style_dim,
-                    use_noise,
-                    demodulate,
-                    img_channels,
-                )
-            )
-
-    def forward(self, images_in, masks_in, ws, noise_mode="random"):
-        x = torch.cat([masks_in - 0.5, images_in * masks_in], dim=1)
-
-        skips = []
-        x, mask = self.conv_first(x, masks_in)  # input size
-        skips.append(x)
-        for i, block in enumerate(self.enc_conv):  # input size to 64
-            x, mask = block(x, mask)
-            if i != len(self.enc_conv) - 1:
-                skips.append(x)
-
-        x_size = x.size()[-2:]
-        x = feature2token(x)
-        mask = feature2token(mask)
-        mid = len(self.tran) // 2
-        for i, block in enumerate(self.tran):  # 64 to 16
-            if i < mid:
-                x, x_size, mask = block(x, x_size, mask)
-                skips.append(x)
-            elif i > mid:
-                x, x_size, mask = block(x, x_size, None)
-                x = x + skips[mid - i]
-            else:
-                x, x_size, mask = block(x, x_size, None)
-
-                mul_map = torch.ones_like(x) * 0.5
-                mul_map = F.dropout(mul_map, training=True).to(x.device)
-                ws = self.ws_style(ws[:, -1]).to(x.device)
-                add_n = self.to_square(ws).unsqueeze(1).to(x.device)
-                add_n = (
-                    F.interpolate(
-                        add_n, size=x.size(1), mode="linear", align_corners=False
-                    )
-                    .squeeze(1)
-                    .unsqueeze(-1)
-                ).to(x.device)
-                x = x * mul_map + add_n * (1 - mul_map)
-                gs = self.to_style(
-                    self.down_conv(token2feature(x, x_size)).flatten(start_dim=1)
-                ).to(x.device)
-                style = torch.cat([gs, ws], dim=1)
-
-        x = token2feature(x, x_size).contiguous()
-        img = None
-        for i, block in enumerate(self.dec_conv):
-            x, img = block(
-                x, img, style, skips[len(self.dec_conv) - i - 1], noise_mode=noise_mode
-            )
-
-        # ensemble
-        img = img * (1 - masks_in) + images_in * masks_in
-
-        return img
-
-
-class SynthesisNet(nn.Module):
-    def __init__(
-        self,
-        w_dim,  # Intermediate latent (W) dimensionality.
-        img_resolution,  # Output image resolution.
-        img_channels=3,  # Number of color channels.
-        channel_base=32768,  # Overall multiplier for the number of channels.
-        channel_decay=1.0,
-        channel_max=512,  # Maximum number of channels in any layer.
-        activation="lrelu",  # Activation function: 'relu', 'lrelu', etc.
-        drop_rate=0.5,
-        use_noise=False,
-        demodulate=True,
-    ):
-        super().__init__()
-        resolution_log2 = int(np.log2(img_resolution))
-        assert img_resolution == 2**resolution_log2 and img_resolution >= 4
-
-        self.num_layers = resolution_log2 * 2 - 3 * 2
-        self.img_resolution = img_resolution
-        self.resolution_log2 = resolution_log2
-
-        # first stage
-        self.first_stage = FirstStage(
-            img_channels,
-            img_resolution=img_resolution,
-            w_dim=w_dim,
-            use_noise=False,
-            demodulate=demodulate,
-        )
-
-        # second stage
-        self.enc = Encoder(
-            resolution_log2, img_channels, activation, patch_size=5, channels=16
-        )
-        self.to_square = FullyConnectedLayer(
-            in_features=w_dim, out_features=16 * 16, activation=activation
-        )
-        self.to_style = ToStyle(
-            in_channels=nf(4),
-            out_channels=nf(2) * 2,
-            activation=activation,
-            drop_rate=drop_rate,
-        )
-        style_dim = w_dim + nf(2) * 2
-        self.dec = Decoder(
-            resolution_log2, activation, style_dim, use_noise, demodulate, img_channels
-        )
-
-    def forward(self, images_in, masks_in, ws, noise_mode="random", return_stg1=False):
-        out_stg1 = self.first_stage(images_in, masks_in, ws, noise_mode=noise_mode)
-
-        # encoder
-        x = images_in * masks_in + out_stg1 * (1 - masks_in)
-        x = torch.cat([masks_in - 0.5, x, images_in * masks_in], dim=1)
-        E_features = self.enc(x)
-
-        fea_16 = E_features[4].to(x.device)
-        mul_map = torch.ones_like(fea_16) * 0.5
-        mul_map = F.dropout(mul_map, training=True).to(x.device)
-        add_n = self.to_square(ws[:, 0]).view(-1, 16, 16).unsqueeze(1)
-        add_n = F.interpolate(
-            add_n, size=fea_16.size()[-2:], mode="bilinear", align_corners=False
-        ).to(x.device)
-        fea_16 = fea_16 * mul_map + add_n * (1 - mul_map)
-        E_features[4] = fea_16
-
-        # style
-        gs = self.to_style(fea_16).to(x.device)
-
-        # decoder
-        img = self.dec(fea_16, ws, gs, E_features, noise_mode=noise_mode).to(x.device)
-
-        # ensemble
-        img = img * (1 - masks_in) + images_in * masks_in
-
-        if not return_stg1:
-            return img
-        else:
-            return img, out_stg1
-
-
-class Generator(nn.Module):
-    def __init__(
-        self,
-        z_dim,  # Input latent (Z) dimensionality, 0 = no latent.
-        c_dim,  # Conditioning label (C) dimensionality, 0 = no label.
-        w_dim,  # Intermediate latent (W) dimensionality.
-        img_resolution,  # resolution of generated image
-        img_channels,  # Number of input color channels.
-        synthesis_kwargs={},  # Arguments for SynthesisNetwork.
-        mapping_kwargs={},  # Arguments for MappingNetwork.
-    ):
-        super().__init__()
-        self.z_dim = z_dim
-        self.c_dim = c_dim
-        self.w_dim = w_dim
-        self.img_resolution = img_resolution
-        self.img_channels = img_channels
-
-        self.synthesis = SynthesisNet(
-            w_dim=w_dim,
-            img_resolution=img_resolution,
-            img_channels=img_channels,
-            **synthesis_kwargs,
-        )
-        self.mapping = MappingNet(
-            z_dim=z_dim,
-            c_dim=c_dim,
-            w_dim=w_dim,
-            num_ws=self.synthesis.num_layers,
-            **mapping_kwargs,
-        )
-
-    def forward(
-        self,
-        images_in,
-        masks_in,
-        z,
-        c,
-        truncation_psi=1,
-        truncation_cutoff=None,
-        skip_w_avg_update=False,
-        noise_mode="none",
-        return_stg1=False,
-    ):
-        ws = self.mapping(
-            z,
-            c,
-            truncation_psi=truncation_psi,
-            truncation_cutoff=truncation_cutoff,
-            skip_w_avg_update=skip_w_avg_update,
-        )
-        img = self.synthesis(images_in, masks_in, ws, noise_mode=noise_mode)
-        return img
-
-
-class MAT(nn.Module):
-    def __init__(self, state_dict):
-        super(MAT, self).__init__()
-        self.model_arch = "MAT"
-        self.sub_type = "Inpaint"
-        self.in_nc = 3
-        self.out_nc = 3
-        self.scale = 1
-
-        self.supports_fp16 = False
-        self.supports_bf16 = True
-
-        self.min_size = 512
-        self.pad_mod = 512
-        self.pad_to_square = True
-
-        seed = 240  # pick up a random number
-        random.seed(seed)
-        np.random.seed(seed)
-        torch.manual_seed(seed)
-
-        self.model = Generator(
-            z_dim=512, c_dim=0, w_dim=512, img_resolution=512, img_channels=3
-        )
-        self.z = torch.from_numpy(np.random.randn(1, self.model.z_dim))  # [1., 512]
-        self.label = torch.zeros([1, self.model.c_dim])
-        self.state = {
-            k.replace("synthesis", "model.synthesis").replace(
-                "mapping", "model.mapping"
-            ): v
-            for k, v in state_dict.items()
-        }
-        self.load_state_dict(self.state, strict=False)
-
-    def forward(self, image, mask):
-        """Input images and output images have same size
-        images: [H, W, C] RGB
-        masks: [H, W] mask area == 255
-        return: BGR IMAGE
-        """
-
-        image = image * 2 - 1  # [0, 1] -> [-1, 1]
-        mask = 1 - mask
-
-        output = self.model(
-            image, mask, self.z, self.label, truncation_psi=1, noise_mode="none"
-        )
-
-        return output * 0.5 + 0.5
diff --git a/comfy_extras/chainner_models/architecture/OmniSR/OmniSR.py b/comfy_extras/chainner_models/architecture/OmniSR/OmniSR.py
index dec16952..1e1c3f35 100644
--- a/comfy_extras/chainner_models/architecture/OmniSR/OmniSR.py
+++ b/comfy_extras/chainner_models/architecture/OmniSR/OmniSR.py
@@ -56,7 +56,17 @@ class OmniSR(nn.Module):
         residual_layer = []
         self.res_num = res_num
 
-        self.window_size = 8  # we can just assume this for now, but there's probably a way to calculate it (just need to get the sqrt of the right layer)
+        if (
+            "residual_layer.0.residual_layer.0.layer.2.fn.rel_pos_bias.weight"
+            in state_dict.keys()
+        ):
+            rel_pos_bias_weight = state_dict[
+                "residual_layer.0.residual_layer.0.layer.2.fn.rel_pos_bias.weight"
+            ].shape[0]
+            self.window_size = int((math.sqrt(rel_pos_bias_weight) + 1) / 2)
+        else:
+            self.window_size = 8
+
         self.up_scale = up_scale
 
         for _ in range(res_num):
diff --git a/comfy_extras/chainner_models/architecture/SCUNet.py b/comfy_extras/chainner_models/architecture/SCUNet.py
new file mode 100644
index 00000000..b8354a87
--- /dev/null
+++ b/comfy_extras/chainner_models/architecture/SCUNet.py
@@ -0,0 +1,455 @@
+# pylint: skip-file
+# -----------------------------------------------------------------------------------
+# SCUNet: Practical Blind Denoising via Swin-Conv-UNet and Data Synthesis, https://arxiv.org/abs/2203.13278
+# Zhang, Kai and Li, Yawei and Liang, Jingyun and Cao, Jiezhang and Zhang, Yulun and Tang, Hao and Timofte, Radu and Van Gool, Luc
+# -----------------------------------------------------------------------------------
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+from .timm.drop import DropPath
+from .timm.weight_init import trunc_normal_
+
+
+# Borrowed from https://github.com/cszn/SCUNet/blob/main/models/network_scunet.py
+class WMSA(nn.Module):
+    """Self-attention module in Swin Transformer"""
+
+    def __init__(self, input_dim, output_dim, head_dim, window_size, type):
+        super(WMSA, self).__init__()
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.head_dim = head_dim
+        self.scale = self.head_dim**-0.5
+        self.n_heads = input_dim // head_dim
+        self.window_size = window_size
+        self.type = type
+        self.embedding_layer = nn.Linear(self.input_dim, 3 * self.input_dim, bias=True)
+
+        self.relative_position_params = nn.Parameter(
+            torch.zeros((2 * window_size - 1) * (2 * window_size - 1), self.n_heads)
+        )
+        # TODO recover
+        # self.relative_position_params = nn.Parameter(torch.zeros(self.n_heads, 2 * window_size - 1, 2 * window_size -1))
+        self.relative_position_params = nn.Parameter(
+            torch.zeros((2 * window_size - 1) * (2 * window_size - 1), self.n_heads)
+        )
+
+        self.linear = nn.Linear(self.input_dim, self.output_dim)
+
+        trunc_normal_(self.relative_position_params, std=0.02)
+        self.relative_position_params = torch.nn.Parameter(
+            self.relative_position_params.view(
+                2 * window_size - 1, 2 * window_size - 1, self.n_heads
+            )
+            .transpose(1, 2)
+            .transpose(0, 1)
+        )
+
+    def generate_mask(self, h, w, p, shift):
+        """generating the mask of SW-MSA
+        Args:
+            shift: shift parameters in CyclicShift.
+        Returns:
+            attn_mask: should be (1 1 w p p),
+        """
+        # supporting square.
+        attn_mask = torch.zeros(
+            h,
+            w,
+            p,
+            p,
+            p,
+            p,
+            dtype=torch.bool,
+            device=self.relative_position_params.device,
+        )
+        if self.type == "W":
+            return attn_mask
+
+        s = p - shift
+        attn_mask[-1, :, :s, :, s:, :] = True
+        attn_mask[-1, :, s:, :, :s, :] = True
+        attn_mask[:, -1, :, :s, :, s:] = True
+        attn_mask[:, -1, :, s:, :, :s] = True
+        attn_mask = rearrange(
+            attn_mask, "w1 w2 p1 p2 p3 p4 -> 1 1 (w1 w2) (p1 p2) (p3 p4)"
+        )
+        return attn_mask
+
+    def forward(self, x):
+        """Forward pass of Window Multi-head Self-attention module.
+        Args:
+            x: input tensor with shape of [b h w c];
+            attn_mask: attention mask, fill -inf where the value is True;
+        Returns:
+            output: tensor shape [b h w c]
+        """
+        if self.type != "W":
+            x = torch.roll(
+                x,
+                shifts=(-(self.window_size // 2), -(self.window_size // 2)),
+                dims=(1, 2),
+            )
+
+        x = rearrange(
+            x,
+            "b (w1 p1) (w2 p2) c -> b w1 w2 p1 p2 c",
+            p1=self.window_size,
+            p2=self.window_size,
+        )
+        h_windows = x.size(1)
+        w_windows = x.size(2)
+        # square validation
+        # assert h_windows == w_windows
+
+        x = rearrange(
+            x,
+            "b w1 w2 p1 p2 c -> b (w1 w2) (p1 p2) c",
+            p1=self.window_size,
+            p2=self.window_size,
+        )
+        qkv = self.embedding_layer(x)
+        q, k, v = rearrange(
+            qkv, "b nw np (threeh c) -> threeh b nw np c", c=self.head_dim
+        ).chunk(3, dim=0)
+        sim = torch.einsum("hbwpc,hbwqc->hbwpq", q, k) * self.scale
+        # Adding learnable relative embedding
+        sim = sim + rearrange(self.relative_embedding(), "h p q -> h 1 1 p q")
+        # Using Attn Mask to distinguish different subwindows.
+        if self.type != "W":
+            attn_mask = self.generate_mask(
+                h_windows, w_windows, self.window_size, shift=self.window_size // 2
+            )
+            sim = sim.masked_fill_(attn_mask, float("-inf"))
+
+        probs = nn.functional.softmax(sim, dim=-1)
+        output = torch.einsum("hbwij,hbwjc->hbwic", probs, v)
+        output = rearrange(output, "h b w p c -> b w p (h c)")
+        output = self.linear(output)
+        output = rearrange(
+            output,
+            "b (w1 w2) (p1 p2) c -> b (w1 p1) (w2 p2) c",
+            w1=h_windows,
+            p1=self.window_size,
+        )
+
+        if self.type != "W":
+            output = torch.roll(
+                output,
+                shifts=(self.window_size // 2, self.window_size // 2),
+                dims=(1, 2),
+            )
+
+        return output
+
+    def relative_embedding(self):
+        cord = torch.tensor(
+            np.array(
+                [
+                    [i, j]
+                    for i in range(self.window_size)
+                    for j in range(self.window_size)
+                ]
+            )
+        )
+        relation = cord[:, None, :] - cord[None, :, :] + self.window_size - 1
+        # negative is allowed
+        return self.relative_position_params[
+            :, relation[:, :, 0].long(), relation[:, :, 1].long()
+        ]
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        input_dim,
+        output_dim,
+        head_dim,
+        window_size,
+        drop_path,
+        type="W",
+        input_resolution=None,
+    ):
+        """SwinTransformer Block"""
+        super(Block, self).__init__()
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        assert type in ["W", "SW"]
+        self.type = type
+        if input_resolution <= window_size:
+            self.type = "W"
+
+        self.ln1 = nn.LayerNorm(input_dim)
+        self.msa = WMSA(input_dim, input_dim, head_dim, window_size, self.type)
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.ln2 = nn.LayerNorm(input_dim)
+        self.mlp = nn.Sequential(
+            nn.Linear(input_dim, 4 * input_dim),
+            nn.GELU(),
+            nn.Linear(4 * input_dim, output_dim),
+        )
+
+    def forward(self, x):
+        x = x + self.drop_path(self.msa(self.ln1(x)))
+        x = x + self.drop_path(self.mlp(self.ln2(x)))
+        return x
+
+
+class ConvTransBlock(nn.Module):
+    def __init__(
+        self,
+        conv_dim,
+        trans_dim,
+        head_dim,
+        window_size,
+        drop_path,
+        type="W",
+        input_resolution=None,
+    ):
+        """SwinTransformer and Conv Block"""
+        super(ConvTransBlock, self).__init__()
+        self.conv_dim = conv_dim
+        self.trans_dim = trans_dim
+        self.head_dim = head_dim
+        self.window_size = window_size
+        self.drop_path = drop_path
+        self.type = type
+        self.input_resolution = input_resolution
+
+        assert self.type in ["W", "SW"]
+        if self.input_resolution <= self.window_size:
+            self.type = "W"
+
+        self.trans_block = Block(
+            self.trans_dim,
+            self.trans_dim,
+            self.head_dim,
+            self.window_size,
+            self.drop_path,
+            self.type,
+            self.input_resolution,
+        )
+        self.conv1_1 = nn.Conv2d(
+            self.conv_dim + self.trans_dim,
+            self.conv_dim + self.trans_dim,
+            1,
+            1,
+            0,
+            bias=True,
+        )
+        self.conv1_2 = nn.Conv2d(
+            self.conv_dim + self.trans_dim,
+            self.conv_dim + self.trans_dim,
+            1,
+            1,
+            0,
+            bias=True,
+        )
+
+        self.conv_block = nn.Sequential(
+            nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False),
+            nn.ReLU(True),
+            nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False),
+        )
+
+    def forward(self, x):
+        conv_x, trans_x = torch.split(
+            self.conv1_1(x), (self.conv_dim, self.trans_dim), dim=1
+        )
+        conv_x = self.conv_block(conv_x) + conv_x
+        trans_x = Rearrange("b c h w -> b h w c")(trans_x)
+        trans_x = self.trans_block(trans_x)
+        trans_x = Rearrange("b h w c -> b c h w")(trans_x)
+        res = self.conv1_2(torch.cat((conv_x, trans_x), dim=1))
+        x = x + res
+
+        return x
+
+
+class SCUNet(nn.Module):
+    def __init__(
+        self,
+        state_dict,
+        in_nc=3,
+        config=[4, 4, 4, 4, 4, 4, 4],
+        dim=64,
+        drop_path_rate=0.0,
+        input_resolution=256,
+    ):
+        super(SCUNet, self).__init__()
+        self.model_arch = "SCUNet"
+        self.sub_type = "SR"
+
+        self.num_filters: int = 0
+
+        self.state = state_dict
+        self.config = config
+        self.dim = dim
+        self.head_dim = 32
+        self.window_size = 8
+
+        self.in_nc = in_nc
+        self.out_nc = self.in_nc
+        self.scale = 1
+        self.supports_fp16 = True
+
+        # drop path rate for each layer
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(config))]
+
+        self.m_head = [nn.Conv2d(in_nc, dim, 3, 1, 1, bias=False)]
+
+        begin = 0
+        self.m_down1 = [
+            ConvTransBlock(
+                dim // 2,
+                dim // 2,
+                self.head_dim,
+                self.window_size,
+                dpr[i + begin],
+                "W" if not i % 2 else "SW",
+                input_resolution,
+            )
+            for i in range(config[0])
+        ] + [nn.Conv2d(dim, 2 * dim, 2, 2, 0, bias=False)]
+
+        begin += config[0]
+        self.m_down2 = [
+            ConvTransBlock(
+                dim,
+                dim,
+                self.head_dim,
+                self.window_size,
+                dpr[i + begin],
+                "W" if not i % 2 else "SW",
+                input_resolution // 2,
+            )
+            for i in range(config[1])
+        ] + [nn.Conv2d(2 * dim, 4 * dim, 2, 2, 0, bias=False)]
+
+        begin += config[1]
+        self.m_down3 = [
+            ConvTransBlock(
+                2 * dim,
+                2 * dim,
+                self.head_dim,
+                self.window_size,
+                dpr[i + begin],
+                "W" if not i % 2 else "SW",
+                input_resolution // 4,
+            )
+            for i in range(config[2])
+        ] + [nn.Conv2d(4 * dim, 8 * dim, 2, 2, 0, bias=False)]
+
+        begin += config[2]
+        self.m_body = [
+            ConvTransBlock(
+                4 * dim,
+                4 * dim,
+                self.head_dim,
+                self.window_size,
+                dpr[i + begin],
+                "W" if not i % 2 else "SW",
+                input_resolution // 8,
+            )
+            for i in range(config[3])
+        ]
+
+        begin += config[3]
+        self.m_up3 = [
+            nn.ConvTranspose2d(8 * dim, 4 * dim, 2, 2, 0, bias=False),
+        ] + [
+            ConvTransBlock(
+                2 * dim,
+                2 * dim,
+                self.head_dim,
+                self.window_size,
+                dpr[i + begin],
+                "W" if not i % 2 else "SW",
+                input_resolution // 4,
+            )
+            for i in range(config[4])
+        ]
+
+        begin += config[4]
+        self.m_up2 = [
+            nn.ConvTranspose2d(4 * dim, 2 * dim, 2, 2, 0, bias=False),
+        ] + [
+            ConvTransBlock(
+                dim,
+                dim,
+                self.head_dim,
+                self.window_size,
+                dpr[i + begin],
+                "W" if not i % 2 else "SW",
+                input_resolution // 2,
+            )
+            for i in range(config[5])
+        ]
+
+        begin += config[5]
+        self.m_up1 = [
+            nn.ConvTranspose2d(2 * dim, dim, 2, 2, 0, bias=False),
+        ] + [
+            ConvTransBlock(
+                dim // 2,
+                dim // 2,
+                self.head_dim,
+                self.window_size,
+                dpr[i + begin],
+                "W" if not i % 2 else "SW",
+                input_resolution,
+            )
+            for i in range(config[6])
+        ]
+
+        self.m_tail = [nn.Conv2d(dim, in_nc, 3, 1, 1, bias=False)]
+
+        self.m_head = nn.Sequential(*self.m_head)
+        self.m_down1 = nn.Sequential(*self.m_down1)
+        self.m_down2 = nn.Sequential(*self.m_down2)
+        self.m_down3 = nn.Sequential(*self.m_down3)
+        self.m_body = nn.Sequential(*self.m_body)
+        self.m_up3 = nn.Sequential(*self.m_up3)
+        self.m_up2 = nn.Sequential(*self.m_up2)
+        self.m_up1 = nn.Sequential(*self.m_up1)
+        self.m_tail = nn.Sequential(*self.m_tail)
+        # self.apply(self._init_weights)
+        self.load_state_dict(state_dict, strict=True)
+
+    def check_image_size(self, x):
+        _, _, h, w = x.size()
+        mod_pad_h = (64 - h % 64) % 64
+        mod_pad_w = (64 - w % 64) % 64
+        x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect")
+        return x
+
+    def forward(self, x0):
+        h, w = x0.size()[-2:]
+        x0 = self.check_image_size(x0)
+
+        x1 = self.m_head(x0)
+        x2 = self.m_down1(x1)
+        x3 = self.m_down2(x2)
+        x4 = self.m_down3(x3)
+        x = self.m_body(x4)
+        x = self.m_up3(x + x4)
+        x = self.m_up2(x + x3)
+        x = self.m_up1(x + x2)
+        x = self.m_tail(x + x1)
+
+        x = x[:, :, :h, :w]
+        return x
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
diff --git a/comfy_extras/chainner_models/architecture/SPSR.py b/comfy_extras/chainner_models/architecture/SPSR.py
index 6f5ac458..c3cefff1 100644
--- a/comfy_extras/chainner_models/architecture/SPSR.py
+++ b/comfy_extras/chainner_models/architecture/SPSR.py
@@ -60,7 +60,6 @@ class SPSRNet(nn.Module):
         self.out_nc: int = self.state["f_HR_conv1.0.bias"].shape[0]
 
         self.scale = self.get_scale(4)
-        print(self.scale)
         self.num_filters: int = self.state["model.0.weight"].shape[0]
 
         self.supports_fp16 = True
diff --git a/comfy_extras/chainner_models/architecture/SwinIR.py b/comfy_extras/chainner_models/architecture/SwinIR.py
index 8cce2d0e..1abf450b 100644
--- a/comfy_extras/chainner_models/architecture/SwinIR.py
+++ b/comfy_extras/chainner_models/architecture/SwinIR.py
@@ -972,6 +972,7 @@ class SwinIR(nn.Module):
         self.upsampler = upsampler
         self.img_size = img_size
         self.img_range = img_range
+        self.resi_connection = resi_connection
 
         self.supports_fp16 = False  # Too much weirdness to support this at the moment
         self.supports_bfp16 = True
diff --git a/comfy_extras/chainner_models/architecture/mat/utils.py b/comfy_extras/chainner_models/architecture/mat/utils.py
deleted file mode 100644
index 1e9445a2..00000000
--- a/comfy_extras/chainner_models/architecture/mat/utils.py
+++ /dev/null
@@ -1,698 +0,0 @@
-"""Code used for this implementation of the MAT helper utils is modified from
-lama-cleaner, copyright of Sanster: https://github.com/fenglinglwb/MAT"""
-
-import collections
-from itertools import repeat
-from typing import Any
-
-import numpy as np
-import torch
-from torch import conv2d, conv_transpose2d
-
-
-def normalize_2nd_moment(x, dim=1, eps=1e-8):
-    return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()
-
-
-class EasyDict(dict):
-    """Convenience class that behaves like a dict but allows access with the attribute syntax."""
-
-    def __getattr__(self, name: str) -> Any:
-        try:
-            return self[name]
-        except KeyError:
-            raise AttributeError(name)
-
-    def __setattr__(self, name: str, value: Any) -> None:
-        self[name] = value
-
-    def __delattr__(self, name: str) -> None:
-        del self[name]
-
-
-activation_funcs = {
-    "linear": EasyDict(
-        func=lambda x, **_: x,
-        def_alpha=0,
-        def_gain=1,
-        cuda_idx=1,
-        ref="",
-        has_2nd_grad=False,
-    ),
-    "relu": EasyDict(
-        func=lambda x, **_: torch.nn.functional.relu(x),
-        def_alpha=0,
-        def_gain=np.sqrt(2),
-        cuda_idx=2,
-        ref="y",
-        has_2nd_grad=False,
-    ),
-    "lrelu": EasyDict(
-        func=lambda x, alpha, **_: torch.nn.functional.leaky_relu(x, alpha),
-        def_alpha=0.2,
-        def_gain=np.sqrt(2),
-        cuda_idx=3,
-        ref="y",
-        has_2nd_grad=False,
-    ),
-    "tanh": EasyDict(
-        func=lambda x, **_: torch.tanh(x),
-        def_alpha=0,
-        def_gain=1,
-        cuda_idx=4,
-        ref="y",
-        has_2nd_grad=True,
-    ),
-    "sigmoid": EasyDict(
-        func=lambda x, **_: torch.sigmoid(x),
-        def_alpha=0,
-        def_gain=1,
-        cuda_idx=5,
-        ref="y",
-        has_2nd_grad=True,
-    ),
-    "elu": EasyDict(
-        func=lambda x, **_: torch.nn.functional.elu(x),
-        def_alpha=0,
-        def_gain=1,
-        cuda_idx=6,
-        ref="y",
-        has_2nd_grad=True,
-    ),
-    "selu": EasyDict(
-        func=lambda x, **_: torch.nn.functional.selu(x),
-        def_alpha=0,
-        def_gain=1,
-        cuda_idx=7,
-        ref="y",
-        has_2nd_grad=True,
-    ),
-    "softplus": EasyDict(
-        func=lambda x, **_: torch.nn.functional.softplus(x),
-        def_alpha=0,
-        def_gain=1,
-        cuda_idx=8,
-        ref="y",
-        has_2nd_grad=True,
-    ),
-    "swish": EasyDict(
-        func=lambda x, **_: torch.sigmoid(x) * x,
-        def_alpha=0,
-        def_gain=np.sqrt(2),
-        cuda_idx=9,
-        ref="x",
-        has_2nd_grad=True,
-    ),
-}
-
-
-def _bias_act_ref(x, b=None, dim=1, act="linear", alpha=None, gain=None, clamp=None):
-    """Slow reference implementation of `bias_act()` using standard TensorFlow ops."""
-    assert isinstance(x, torch.Tensor)
-    assert clamp is None or clamp >= 0
-    spec = activation_funcs[act]
-    alpha = float(alpha if alpha is not None else spec.def_alpha)
-    gain = float(gain if gain is not None else spec.def_gain)
-    clamp = float(clamp if clamp is not None else -1)
-
-    # Add bias.
-    if b is not None:
-        assert isinstance(b, torch.Tensor) and b.ndim == 1
-        assert 0 <= dim < x.ndim
-        assert b.shape[0] == x.shape[dim]
-        x = x + b.reshape([-1 if i == dim else 1 for i in range(x.ndim)]).to(x.device)
-
-    # Evaluate activation function.
-    alpha = float(alpha)
-    x = spec.func(x, alpha=alpha)
-
-    # Scale by gain.
-    gain = float(gain)
-    if gain != 1:
-        x = x * gain
-
-    # Clamp.
-    if clamp >= 0:
-        x = x.clamp(-clamp, clamp)  # pylint: disable=invalid-unary-operand-type
-    return x
-
-
-def bias_act(
-    x, b=None, dim=1, act="linear", alpha=None, gain=None, clamp=None, impl="ref"
-):
-    r"""Fused bias and activation function.
-    Adds bias `b` to activation tensor `x`, evaluates activation function `act`,
-    and scales the result by `gain`. Each of the steps is optional. In most cases,
-    the fused op is considerably more efficient than performing the same calculation
-    using standard PyTorch ops. It supports first and second order gradients,
-    but not third order gradients.
-    Args:
-        x:      Input activation tensor. Can be of any shape.
-        b:      Bias vector, or `None` to disable. Must be a 1D tensor of the same type
-                as `x`. The shape must be known, and it must match the dimension of `x`
-                corresponding to `dim`.
-        dim:    The dimension in `x` corresponding to the elements of `b`.
-                The value of `dim` is ignored if `b` is not specified.
-        act:    Name of the activation function to evaluate, or `"linear"` to disable.
-                Can be e.g. `"relu"`, `"lrelu"`, `"tanh"`, `"sigmoid"`, `"swish"`, etc.
-                See `activation_funcs` for a full list. `None` is not allowed.
-        alpha:  Shape parameter for the activation function, or `None` to use the default.
-        gain:   Scaling factor for the output tensor, or `None` to use default.
-                See `activation_funcs` for the default scaling of each activation function.
-                If unsure, consider specifying 1.
-        clamp:  Clamp the output values to `[-clamp, +clamp]`, or `None` to disable
-                the clamping (default).
-        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
-    Returns:
-        Tensor of the same shape and datatype as `x`.
-    """
-    assert isinstance(x, torch.Tensor)
-    assert impl in ["ref", "cuda"]
-    return _bias_act_ref(
-        x=x, b=b, dim=dim, act=act, alpha=alpha, gain=gain, clamp=clamp
-    )
-
-
-def setup_filter(
-    f,
-    device=torch.device("cpu"),
-    normalize=True,
-    flip_filter=False,
-    gain=1,
-    separable=None,
-):
-    r"""Convenience function to setup 2D FIR filter for `upfirdn2d()`.
-    Args:
-        f:           Torch tensor, numpy array, or python list of the shape
-                     `[filter_height, filter_width]` (non-separable),
-                     `[filter_taps]` (separable),
-                     `[]` (impulse), or
-                     `None` (identity).
-        device:      Result device (default: cpu).
-        normalize:   Normalize the filter so that it retains the magnitude
-                     for constant input signal (DC)? (default: True).
-        flip_filter: Flip the filter? (default: False).
-        gain:        Overall scaling factor for signal magnitude (default: 1).
-        separable:   Return a separable filter? (default: select automatically).
-    Returns:
-        Float32 tensor of the shape
-        `[filter_height, filter_width]` (non-separable) or
-        `[filter_taps]` (separable).
-    """
-    # Validate.
-    if f is None:
-        f = 1
-    f = torch.as_tensor(f, dtype=torch.float32)
-    assert f.ndim in [0, 1, 2]
-    assert f.numel() > 0
-    if f.ndim == 0:
-        f = f[np.newaxis]
-
-    # Separable?
-    if separable is None:
-        separable = f.ndim == 1 and f.numel() >= 8
-    if f.ndim == 1 and not separable:
-        f = f.ger(f)
-    assert f.ndim == (1 if separable else 2)
-
-    # Apply normalize, flip, gain, and device.
-    if normalize:
-        f /= f.sum()
-    if flip_filter:
-        f = f.flip(list(range(f.ndim)))
-    f = f * (gain ** (f.ndim / 2))
-    f = f.to(device=device)
-    return f
-
-
-def _get_filter_size(f):
-    if f is None:
-        return 1, 1
-
-    assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
-    fw = f.shape[-1]
-    fh = f.shape[0]
-
-    fw = int(fw)
-    fh = int(fh)
-    assert fw >= 1 and fh >= 1
-    return fw, fh
-
-
-def _get_weight_shape(w):
-    shape = [int(sz) for sz in w.shape]
-    return shape
-
-
-def _parse_scaling(scaling):
-    if isinstance(scaling, int):
-        scaling = [scaling, scaling]
-    assert isinstance(scaling, (list, tuple))
-    assert all(isinstance(x, int) for x in scaling)
-    sx, sy = scaling
-    assert sx >= 1 and sy >= 1
-    return sx, sy
-
-
-def _parse_padding(padding):
-    if isinstance(padding, int):
-        padding = [padding, padding]
-    assert isinstance(padding, (list, tuple))
-    assert all(isinstance(x, int) for x in padding)
-    if len(padding) == 2:
-        padx, pady = padding
-        padding = [padx, padx, pady, pady]
-    padx0, padx1, pady0, pady1 = padding
-    return padx0, padx1, pady0, pady1
-
-
-def _ntuple(n):
-    def parse(x):
-        if isinstance(x, collections.abc.Iterable):
-            return x
-        return tuple(repeat(x, n))
-
-    return parse
-
-
-to_2tuple = _ntuple(2)
-
-
-def _upfirdn2d_ref(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1):
-    """Slow reference implementation of `upfirdn2d()` using standard PyTorch ops."""
-    # Validate arguments.
-    assert isinstance(x, torch.Tensor) and x.ndim == 4
-    if f is None:
-        f = torch.ones([1, 1], dtype=torch.float32, device=x.device)
-    assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
-    assert f.dtype == torch.float32 and not f.requires_grad
-    batch_size, num_channels, in_height, in_width = x.shape
-    # upx, upy = _parse_scaling(up)
-    # downx, downy = _parse_scaling(down)
-
-    upx, upy = up, up
-    downx, downy = down, down
-
-    # padx0, padx1, pady0, pady1 = _parse_padding(padding)
-    padx0, padx1, pady0, pady1 = padding[0], padding[1], padding[2], padding[3]
-
-    # Upsample by inserting zeros.
-    x = x.reshape([batch_size, num_channels, in_height, 1, in_width, 1])
-    x = torch.nn.functional.pad(x, [0, upx - 1, 0, 0, 0, upy - 1])
-    x = x.reshape([batch_size, num_channels, in_height * upy, in_width * upx])
-
-    # Pad or crop.
-    x = torch.nn.functional.pad(
-        x, [max(padx0, 0), max(padx1, 0), max(pady0, 0), max(pady1, 0)]
-    )
-    x = x[
-        :,
-        :,
-        max(-pady0, 0) : x.shape[2] - max(-pady1, 0),
-        max(-padx0, 0) : x.shape[3] - max(-padx1, 0),
-    ]
-
-    # Setup filter.
-    f = f * (gain ** (f.ndim / 2))
-    f = f.to(x.dtype)
-    if not flip_filter:
-        f = f.flip(list(range(f.ndim)))
-
-    # Convolve with the filter.
-    f = f[np.newaxis, np.newaxis].repeat([num_channels, 1] + [1] * f.ndim)
-    if f.ndim == 4:
-        x = conv2d(input=x, weight=f, groups=num_channels)
-    else:
-        x = conv2d(input=x, weight=f.unsqueeze(2), groups=num_channels)
-        x = conv2d(input=x, weight=f.unsqueeze(3), groups=num_channels)
-
-    # Downsample by throwing away pixels.
-    x = x[:, :, ::downy, ::downx]
-    return x
-
-
-def upfirdn2d(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1, impl="cuda"):
-    r"""Pad, upsample, filter, and downsample a batch of 2D images.
-    Performs the following sequence of operations for each channel:
-    1. Upsample the image by inserting N-1 zeros after each pixel (`up`).
-    2. Pad the image with the specified number of zeros on each side (`padding`).
-       Negative padding corresponds to cropping the image.
-    3. Convolve the image with the specified 2D FIR filter (`f`), shrinking it
-       so that the footprint of all output pixels lies within the input image.
-    4. Downsample the image by keeping every Nth pixel (`down`).
-    This sequence of operations bears close resemblance to scipy.signal.upfirdn().
-    The fused op is considerably more efficient than performing the same calculation
-    using standard PyTorch ops. It supports gradients of arbitrary order.
-    Args:
-        x:           Float32/float64/float16 input tensor of the shape
-                     `[batch_size, num_channels, in_height, in_width]`.
-        f:           Float32 FIR filter of the shape
-                     `[filter_height, filter_width]` (non-separable),
-                     `[filter_taps]` (separable), or
-                     `None` (identity).
-        up:          Integer upsampling factor. Can be a single int or a list/tuple
-                     `[x, y]` (default: 1).
-        down:        Integer downsampling factor. Can be a single int or a list/tuple
-                     `[x, y]` (default: 1).
-        padding:     Padding with respect to the upsampled image. Can be a single number
-                     or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
-                     (default: 0).
-        flip_filter: False = convolution, True = correlation (default: False).
-        gain:        Overall scaling factor for signal magnitude (default: 1).
-        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
-    Returns:
-        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
-    """
-    # assert isinstance(x, torch.Tensor)
-    # assert impl in ['ref', 'cuda']
-    return _upfirdn2d_ref(
-        x, f, up=up, down=down, padding=padding, flip_filter=flip_filter, gain=gain
-    )
-
-
-def upsample2d(x, f, up=2, padding=0, flip_filter=False, gain=1, impl="cuda"):
-    r"""Upsample a batch of 2D images using the given 2D FIR filter.
-    By default, the result is padded so that its shape is a multiple of the input.
-    User-specified padding is applied on top of that, with negative values
-    indicating cropping. Pixels outside the image are assumed to be zero.
-    Args:
-        x:           Float32/float64/float16 input tensor of the shape
-                     `[batch_size, num_channels, in_height, in_width]`.
-        f:           Float32 FIR filter of the shape
-                     `[filter_height, filter_width]` (non-separable),
-                     `[filter_taps]` (separable), or
-                     `None` (identity).
-        up:          Integer upsampling factor. Can be a single int or a list/tuple
-                     `[x, y]` (default: 1).
-        padding:     Padding with respect to the output. Can be a single number or a
-                     list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
-                     (default: 0).
-        flip_filter: False = convolution, True = correlation (default: False).
-        gain:        Overall scaling factor for signal magnitude (default: 1).
-        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
-    Returns:
-        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
-    """
-    upx, upy = _parse_scaling(up)
-    # upx, upy = up, up
-    padx0, padx1, pady0, pady1 = _parse_padding(padding)
-    # padx0, padx1, pady0, pady1 = padding, padding, padding, padding
-    fw, fh = _get_filter_size(f)
-    p = [
-        padx0 + (fw + upx - 1) // 2,
-        padx1 + (fw - upx) // 2,
-        pady0 + (fh + upy - 1) // 2,
-        pady1 + (fh - upy) // 2,
-    ]
-    return upfirdn2d(
-        x,
-        f,
-        up=up,
-        padding=p,
-        flip_filter=flip_filter,
-        gain=gain * upx * upy,
-        impl=impl,
-    )
-
-
-class FullyConnectedLayer(torch.nn.Module):
-    def __init__(
-        self,
-        in_features,  # Number of input features.
-        out_features,  # Number of output features.
-        bias=True,  # Apply additive bias before the activation function?
-        activation="linear",  # Activation function: 'relu', 'lrelu', etc.
-        lr_multiplier=1,  # Learning rate multiplier.
-        bias_init=0,  # Initial value for the additive bias.
-    ):
-        super().__init__()
-        self.weight = torch.nn.Parameter(
-            torch.randn([out_features, in_features]) / lr_multiplier
-        )
-        self.bias = (
-            torch.nn.Parameter(torch.full([out_features], np.float32(bias_init)))
-            if bias
-            else None
-        )
-        self.activation = activation
-
-        self.weight_gain = lr_multiplier / np.sqrt(in_features)
-        self.bias_gain = lr_multiplier
-
-    def forward(self, x):
-        w = self.weight * self.weight_gain
-        b = self.bias
-        if b is not None and self.bias_gain != 1:
-            b = b * self.bias_gain
-
-        if self.activation == "linear" and b is not None:
-            # out = torch.addmm(b.unsqueeze(0), x, w.t())
-            x = x.matmul(w.t().to(x.device))
-            out = x + b.reshape(
-                [-1 if i == x.ndim - 1 else 1 for i in range(x.ndim)]
-            ).to(x.device)
-        else:
-            x = x.matmul(w.t().to(x.device))
-            out = bias_act(x, b, act=self.activation, dim=x.ndim - 1).to(x.device)
-        return out
-
-
-def _conv2d_wrapper(
-    x, w, stride=1, padding=0, groups=1, transpose=False, flip_weight=True
-):
-    """Wrapper for the underlying `conv2d()` and `conv_transpose2d()` implementations."""
-    out_channels, in_channels_per_group, kh, kw = _get_weight_shape(w)
-
-    # Flip weight if requested.
-    if (
-        not flip_weight
-    ):  # conv2d() actually performs correlation (flip_weight=True) not convolution (flip_weight=False).
-        w = w.flip([2, 3])
-
-    # Workaround performance pitfall in cuDNN 8.0.5, triggered when using
-    # 1x1 kernel + memory_format=channels_last + less than 64 channels.
-    if (
-        kw == 1
-        and kh == 1
-        and stride == 1
-        and padding in [0, [0, 0], (0, 0)]
-        and not transpose
-    ):
-        if x.stride()[1] == 1 and min(out_channels, in_channels_per_group) < 64:
-            if out_channels <= 4 and groups == 1:
-                in_shape = x.shape
-                x = w.squeeze(3).squeeze(2) @ x.reshape(
-                    [in_shape[0], in_channels_per_group, -1]
-                )
-                x = x.reshape([in_shape[0], out_channels, in_shape[2], in_shape[3]])
-            else:
-                x = x.to(memory_format=torch.contiguous_format)
-                w = w.to(memory_format=torch.contiguous_format)
-                x = conv2d(x, w, groups=groups)
-            return x.to(memory_format=torch.channels_last)
-
-    # Otherwise => execute using conv2d_gradfix.
-    op = conv_transpose2d if transpose else conv2d
-    return op(x, w, stride=stride, padding=padding, groups=groups)
-
-
-def conv2d_resample(
-    x, w, f=None, up=1, down=1, padding=0, groups=1, flip_weight=True, flip_filter=False
-):
-    r"""2D convolution with optional up/downsampling.
-    Padding is performed only once at the beginning, not between the operations.
-    Args:
-        x:              Input tensor of shape
-                        `[batch_size, in_channels, in_height, in_width]`.
-        w:              Weight tensor of shape
-                        `[out_channels, in_channels//groups, kernel_height, kernel_width]`.
-        f:              Low-pass filter for up/downsampling. Must be prepared beforehand by
-                        calling setup_filter(). None = identity (default).
-        up:             Integer upsampling factor (default: 1).
-        down:           Integer downsampling factor (default: 1).
-        padding:        Padding with respect to the upsampled image. Can be a single number
-                        or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
-                        (default: 0).
-        groups:         Split input channels into N groups (default: 1).
-        flip_weight:    False = convolution, True = correlation (default: True).
-        flip_filter:    False = convolution, True = correlation (default: False).
-    Returns:
-        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
-    """
-    # Validate arguments.
-    assert isinstance(x, torch.Tensor) and (x.ndim == 4)
-    assert isinstance(w, torch.Tensor) and (w.ndim == 4) and (w.dtype == x.dtype)
-    assert f is None or (
-        isinstance(f, torch.Tensor) and f.ndim in [1, 2] and f.dtype == torch.float32
-    )
-    assert isinstance(up, int) and (up >= 1)
-    assert isinstance(down, int) and (down >= 1)
-    # assert isinstance(groups, int) and (groups >= 1), f"!!!!!! groups: {groups} isinstance(groups, int)  {isinstance(groups, int)} {type(groups)}"
-    out_channels, in_channels_per_group, kh, kw = _get_weight_shape(w)
-    fw, fh = _get_filter_size(f)
-    # px0, px1, py0, py1 = _parse_padding(padding)
-    px0, px1, py0, py1 = padding, padding, padding, padding
-
-    # Adjust padding to account for up/downsampling.
-    if up > 1:
-        px0 += (fw + up - 1) // 2
-        px1 += (fw - up) // 2
-        py0 += (fh + up - 1) // 2
-        py1 += (fh - up) // 2
-    if down > 1:
-        px0 += (fw - down + 1) // 2
-        px1 += (fw - down) // 2
-        py0 += (fh - down + 1) // 2
-        py1 += (fh - down) // 2
-
-    # Fast path: 1x1 convolution with downsampling only => downsample first, then convolve.
-    if kw == 1 and kh == 1 and (down > 1 and up == 1):
-        x = upfirdn2d(
-            x=x, f=f, down=down, padding=[px0, px1, py0, py1], flip_filter=flip_filter
-        )
-        x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
-        return x
-
-    # Fast path: 1x1 convolution with upsampling only => convolve first, then upsample.
-    if kw == 1 and kh == 1 and (up > 1 and down == 1):
-        x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
-        x = upfirdn2d(
-            x=x,
-            f=f,
-            up=up,
-            padding=[px0, px1, py0, py1],
-            gain=up**2,
-            flip_filter=flip_filter,
-        )
-        return x
-
-    # Fast path: downsampling only => use strided convolution.
-    if down > 1 and up == 1:
-        x = upfirdn2d(x=x, f=f, padding=[px0, px1, py0, py1], flip_filter=flip_filter)
-        x = _conv2d_wrapper(
-            x=x, w=w, stride=down, groups=groups, flip_weight=flip_weight
-        )
-        return x
-
-    # Fast path: upsampling with optional downsampling => use transpose strided convolution.
-    if up > 1:
-        if groups == 1:
-            w = w.transpose(0, 1)
-        else:
-            w = w.reshape(groups, out_channels // groups, in_channels_per_group, kh, kw)
-            w = w.transpose(1, 2)
-            w = w.reshape(
-                groups * in_channels_per_group, out_channels // groups, kh, kw
-            )
-        px0 -= kw - 1
-        px1 -= kw - up
-        py0 -= kh - 1
-        py1 -= kh - up
-        pxt = max(min(-px0, -px1), 0)
-        pyt = max(min(-py0, -py1), 0)
-        x = _conv2d_wrapper(
-            x=x,
-            w=w,
-            stride=up,
-            padding=[pyt, pxt],
-            groups=groups,
-            transpose=True,
-            flip_weight=(not flip_weight),
-        )
-        x = upfirdn2d(
-            x=x,
-            f=f,
-            padding=[px0 + pxt, px1 + pxt, py0 + pyt, py1 + pyt],
-            gain=up**2,
-            flip_filter=flip_filter,
-        )
-        if down > 1:
-            x = upfirdn2d(x=x, f=f, down=down, flip_filter=flip_filter)
-        return x
-
-    # Fast path: no up/downsampling, padding supported by the underlying implementation => use plain conv2d.
-    if up == 1 and down == 1:
-        if px0 == px1 and py0 == py1 and px0 >= 0 and py0 >= 0:
-            return _conv2d_wrapper(
-                x=x, w=w, padding=[py0, px0], groups=groups, flip_weight=flip_weight
-            )
-
-    # Fallback: Generic reference implementation.
-    x = upfirdn2d(
-        x=x,
-        f=(f if up > 1 else None),
-        up=up,
-        padding=[px0, px1, py0, py1],
-        gain=up**2,
-        flip_filter=flip_filter,
-    )
-    x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
-    if down > 1:
-        x = upfirdn2d(x=x, f=f, down=down, flip_filter=flip_filter)
-    return x
-
-
-class Conv2dLayer(torch.nn.Module):
-    def __init__(
-        self,
-        in_channels,  # Number of input channels.
-        out_channels,  # Number of output channels.
-        kernel_size,  # Width and height of the convolution kernel.
-        bias=True,  # Apply additive bias before the activation function?
-        activation="linear",  # Activation function: 'relu', 'lrelu', etc.
-        up=1,  # Integer upsampling factor.
-        down=1,  # Integer downsampling factor.
-        resample_filter=[
-            1,
-            3,
-            3,
-            1,
-        ],  # Low-pass filter to apply when resampling activations.
-        conv_clamp=None,  # Clamp the output to +-X, None = disable clamping.
-        channels_last=False,  # Expect the input to have memory_format=channels_last?
-        trainable=True,  # Update the weights of this layer during training?
-    ):
-        super().__init__()
-        self.activation = activation
-        self.up = up
-        self.down = down
-        self.register_buffer("resample_filter", setup_filter(resample_filter))
-        self.conv_clamp = conv_clamp
-        self.padding = kernel_size // 2
-        self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size**2))
-        self.act_gain = activation_funcs[activation].def_gain
-
-        memory_format = (
-            torch.channels_last if channels_last else torch.contiguous_format
-        )
-        weight = torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(
-            memory_format=memory_format
-        )
-        bias = torch.zeros([out_channels]) if bias else None
-        if trainable:
-            self.weight = torch.nn.Parameter(weight)
-            self.bias = torch.nn.Parameter(bias) if bias is not None else None
-        else:
-            self.register_buffer("weight", weight)
-            if bias is not None:
-                self.register_buffer("bias", bias)
-            else:
-                self.bias = None
-
-    def forward(self, x, gain=1):
-        w = self.weight * self.weight_gain
-        x = conv2d_resample(
-            x=x,
-            w=w,
-            f=self.resample_filter,
-            up=self.up,
-            down=self.down,
-            padding=self.padding,
-        )
-
-        act_gain = self.act_gain * gain
-        act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
-        out = bias_act(
-            x, self.bias, act=self.activation, gain=act_gain, clamp=act_clamp
-        )
-        return out
diff --git a/comfy_extras/chainner_models/model_loading.py b/comfy_extras/chainner_models/model_loading.py
index 2e66e624..e000871c 100644
--- a/comfy_extras/chainner_models/model_loading.py
+++ b/comfy_extras/chainner_models/model_loading.py
@@ -1,13 +1,14 @@
 import logging as logger
 
+from .architecture.DAT import DAT
 from .architecture.face.codeformer import CodeFormer
 from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean
 from .architecture.face.restoreformer_arch import RestoreFormer
 from .architecture.HAT import HAT
 from .architecture.LaMa import LaMa
-from .architecture.MAT import MAT
 from .architecture.OmniSR.OmniSR import OmniSR
 from .architecture.RRDB import RRDBNet as ESRGAN
+from .architecture.SCUNet import SCUNet
 from .architecture.SPSR import SPSRNet as SPSR
 from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2
 from .architecture.SwiftSRGAN import Generator as SwiftSRGAN
@@ -33,7 +34,6 @@ def load_state_dict(state_dict) -> PyTorchModel:
         state_dict = state_dict["params"]
 
     state_dict_keys = list(state_dict.keys())
-
     # SRVGGNet Real-ESRGAN (v2)
     if "body.0.weight" in state_dict_keys and "body.1.weight" in state_dict_keys:
         model = RealESRGANv2(state_dict)
@@ -46,12 +46,14 @@ def load_state_dict(state_dict) -> PyTorchModel:
         and "initial.cnn.depthwise.weight" in state_dict["model"].keys()
     ):
         model = SwiftSRGAN(state_dict)
-    # HAT -- be sure it is above swinir
-    elif "layers.0.residual_group.blocks.0.conv_block.cab.0.weight" in state_dict_keys:
-        model = HAT(state_dict)
-    # SwinIR
+    # SwinIR, Swin2SR, HAT
     elif "layers.0.residual_group.blocks.0.norm1.weight" in state_dict_keys:
-        if "patch_embed.proj.weight" in state_dict_keys:
+        if (
+            "layers.0.residual_group.blocks.0.conv_block.cab.0.weight"
+            in state_dict_keys
+        ):
+            model = HAT(state_dict)
+        elif "patch_embed.proj.weight" in state_dict_keys:
             model = Swin2SR(state_dict)
         else:
             model = SwinIR(state_dict)
@@ -78,12 +80,15 @@ def load_state_dict(state_dict) -> PyTorchModel:
         or "generator.model.1.bn_l.running_mean" in state_dict_keys
     ):
         model = LaMa(state_dict)
-    # MAT
-    elif "synthesis.first_stage.conv_first.conv.resample_filter" in state_dict_keys:
-        model = MAT(state_dict)
     # Omni-SR
     elif "residual_layer.0.residual_layer.0.layer.0.fn.0.weight" in state_dict_keys:
         model = OmniSR(state_dict)
+    # SCUNet
+    elif "m_head.0.weight" in state_dict_keys and "m_tail.0.weight" in state_dict_keys:
+        model = SCUNet(state_dict)
+    # DAT
+    elif "layers.0.blocks.2.attn.attn_mask_0" in state_dict_keys:
+        model = DAT(state_dict)
     # Regular ESRGAN, "new-arch" ESRGAN, Real-ESRGAN v1
     else:
         try:
diff --git a/comfy_extras/chainner_models/types.py b/comfy_extras/chainner_models/types.py
index 1906c0c7..193333b9 100644
--- a/comfy_extras/chainner_models/types.py
+++ b/comfy_extras/chainner_models/types.py
@@ -1,20 +1,32 @@
 from typing import Union
 
+from .architecture.DAT import DAT
 from .architecture.face.codeformer import CodeFormer
 from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean
 from .architecture.face.restoreformer_arch import RestoreFormer
 from .architecture.HAT import HAT
 from .architecture.LaMa import LaMa
-from .architecture.MAT import MAT
 from .architecture.OmniSR.OmniSR import OmniSR
 from .architecture.RRDB import RRDBNet as ESRGAN
+from .architecture.SCUNet import SCUNet
 from .architecture.SPSR import SPSRNet as SPSR
 from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2
 from .architecture.SwiftSRGAN import Generator as SwiftSRGAN
 from .architecture.Swin2SR import Swin2SR
 from .architecture.SwinIR import SwinIR
 
-PyTorchSRModels = (RealESRGANv2, SPSR, SwiftSRGAN, ESRGAN, SwinIR, Swin2SR, HAT, OmniSR)
+PyTorchSRModels = (
+    RealESRGANv2,
+    SPSR,
+    SwiftSRGAN,
+    ESRGAN,
+    SwinIR,
+    Swin2SR,
+    HAT,
+    OmniSR,
+    SCUNet,
+    DAT,
+)
 PyTorchSRModel = Union[
     RealESRGANv2,
     SPSR,
@@ -24,6 +36,8 @@ PyTorchSRModel = Union[
     Swin2SR,
     HAT,
     OmniSR,
+    SCUNet,
+    DAT,
 ]
 
 
@@ -39,8 +53,8 @@ def is_pytorch_face_model(model: object):
     return isinstance(model, PyTorchFaceModels)
 
 
-PyTorchInpaintModels = (LaMa, MAT)
-PyTorchInpaintModel = Union[LaMa, MAT]
+PyTorchInpaintModels = (LaMa,)
+PyTorchInpaintModel = Union[LaMa]
 
 
 def is_pytorch_inpaint_model(model: object):
diff --git a/comfy_extras/nodes_canny.py b/comfy_extras/nodes_canny.py
index d7c3f132..94d453f2 100644
--- a/comfy_extras/nodes_canny.py
+++ b/comfy_extras/nodes_canny.py
@@ -3,7 +3,7 @@ import math
 
 import torch
 import torch.nn.functional as F
-
+import comfy.model_management
 
 def get_canny_nms_kernel(device=None, dtype=None):
     """Utility function that returns 3x3 kernels for the Canny Non-maximal suppression."""
@@ -290,8 +290,8 @@ class Canny:
     CATEGORY = "image/preprocessors"
 
     def detect_edge(self, image, low_threshold, high_threshold):
-        output = canny(image.movedim(-1, 1), low_threshold, high_threshold)
-        img_out = output[1].repeat(1, 3, 1, 1).movedim(1, -1)
+        output = canny(image.to(comfy.model_management.get_torch_device()).movedim(-1, 1), low_threshold, high_threshold)
+        img_out = output[1].cpu().repeat(1, 3, 1, 1).movedim(1, -1)
         return (img_out,)
 
 NODE_CLASS_MAPPINGS = {
diff --git a/comfy_extras/nodes_mask.py b/comfy_extras/nodes_mask.py
index 5adb468a..43f623a6 100644
--- a/comfy_extras/nodes_mask.py
+++ b/comfy_extras/nodes_mask.py
@@ -125,6 +125,27 @@ class ImageToMask:
         mask = image[0, :, :, channels.index(channel)]
         return (mask,)
 
+class ImageColorToMask:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                    "image": ("IMAGE",),
+                    "color": ("INT", {"default": 0, "min": 0, "max": 0xFFFFFF, "step": 1, "display": "color"}),
+                }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+    FUNCTION = "image_to_mask"
+
+    def image_to_mask(self, image, color):
+        temp = (torch.clamp(image[0], 0, 1.0) * 255.0).round().to(torch.int)
+        temp = torch.bitwise_left_shift(temp[:,:,0], 16) + torch.bitwise_left_shift(temp[:,:,1], 8) + temp[:,:,2]
+        mask = torch.where(temp == color, 255, 0).float()
+        return (mask,)
+
 class SolidMask:
     @classmethod
     def INPUT_TYPES(cls):
@@ -315,6 +336,7 @@ NODE_CLASS_MAPPINGS = {
     "ImageCompositeMasked": ImageCompositeMasked,
     "MaskToImage": MaskToImage,
     "ImageToMask": ImageToMask,
+    "ImageColorToMask": ImageColorToMask,
     "SolidMask": SolidMask,
     "InvertMask": InvertMask,
     "CropMask": CropMask,
diff --git a/folder_paths.py b/folder_paths.py
index fa2998fd..be08da18 100644
--- a/folder_paths.py
+++ b/folder_paths.py
@@ -195,17 +195,22 @@ def add_model_folder_path(folder_name, full_folder_path):
 def get_folder_paths(folder_name):
     return folder_names_and_paths[folder_name][0][:]
 
-def recursive_search(directory):
+def recursive_search(directory, excluded_dir_names=None):
     if not os.path.isdir(directory):
         return [], {}
+
+    if excluded_dir_names is None:
+        excluded_dir_names = []
+
     result = []
     dirs = {directory: os.path.getmtime(directory)}
-    for root, subdir, file in os.walk(directory, followlinks=True):
-        for filepath in file:
-            #we os.path,join directory with a blank string to generate a path separator at the end.
-            result.append(os.path.join(root, filepath).replace(os.path.join(directory,''),''))
-        for d in subdir:
-            path = os.path.join(root, d)
+    for dirpath, subdirs, filenames in os.walk(directory, followlinks=True, topdown=True):
+        subdirs[:] = [d for d in subdirs if d not in excluded_dir_names]
+        for file_name in filenames:
+            relative_path = os.path.relpath(os.path.join(dirpath, file_name), directory)
+            result.append(relative_path)
+        for d in subdirs:
+            path = os.path.join(dirpath, d)
             dirs[path] = os.path.getmtime(path)
     return result, dirs
 
@@ -233,7 +238,7 @@ def get_filename_list_(folder_name):
     folders = folder_names_and_paths[folder_name]
     output_folders = {}
     for x in folders[0]:
-        files, folders_all = recursive_search(x)
+        files, folders_all = recursive_search(x, excluded_dir_names=[".git"])
         output_list.update(filter_files_extensions(files, folders[1]))
         output_folders = {**output_folders, **folders_all}
 
diff --git a/nodes.py b/nodes.py
index b2f224ea..77d18052 100644
--- a/nodes.py
+++ b/nodes.py
@@ -22,6 +22,7 @@ import comfy.samplers
 import comfy.sample
 import comfy.sd
 import comfy.utils
+import comfy.controlnet
 
 import comfy.clip_vision
 
@@ -158,6 +159,31 @@ class ConditioningSetArea:
             c.append(n)
         return (c, )
 
+class ConditioningSetAreaPercentage:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"conditioning": ("CONDITIONING", ),
+                              "width": ("FLOAT", {"default": 1.0, "min": 0, "max": 1.0, "step": 0.01}),
+                              "height": ("FLOAT", {"default": 1.0, "min": 0, "max": 1.0, "step": 0.01}),
+                              "x": ("FLOAT", {"default": 0, "min": 0, "max": 1.0, "step": 0.01}),
+                              "y": ("FLOAT", {"default": 0, "min": 0, "max": 1.0, "step": 0.01}),
+                              "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
+                             }}
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "append"
+
+    CATEGORY = "conditioning"
+
+    def append(self, conditioning, width, height, x, y, strength):
+        c = []
+        for t in conditioning:
+            n = [t[0], t[1].copy()]
+            n[1]['area'] = ("percentage", height, width, y, x)
+            n[1]['strength'] = strength
+            n[1]['set_area_to_bounds'] = False
+            c.append(n)
+        return (c, )
+
 class ConditioningSetMask:
     @classmethod
     def INPUT_TYPES(s):
@@ -243,14 +269,16 @@ class VAEDecode:
 class VAEDecodeTiled:
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "samples": ("LATENT", ), "vae": ("VAE", )}}
+        return {"required": {"samples": ("LATENT", ), "vae": ("VAE", ),
+                             "tile_size": ("INT", {"default": 512, "min": 320, "max": 4096, "step": 64})
+                            }}
     RETURN_TYPES = ("IMAGE",)
     FUNCTION = "decode"
 
     CATEGORY = "_for_testing"
 
-    def decode(self, vae, samples):
-        return (vae.decode_tiled(samples["samples"]), )
+    def decode(self, vae, samples, tile_size):
+        return (vae.decode_tiled(samples["samples"], tile_x=tile_size // 8, tile_y=tile_size // 8, ), )
 
 class VAEEncode:
     @classmethod
@@ -279,15 +307,17 @@ class VAEEncode:
 class VAEEncodeTiled:
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "pixels": ("IMAGE", ), "vae": ("VAE", )}}
+        return {"required": {"pixels": ("IMAGE", ), "vae": ("VAE", ),
+                             "tile_size": ("INT", {"default": 512, "min": 320, "max": 4096, "step": 64})
+                            }}
     RETURN_TYPES = ("LATENT",)
     FUNCTION = "encode"
 
     CATEGORY = "_for_testing"
 
-    def encode(self, vae, pixels):
+    def encode(self, vae, pixels, tile_size):
         pixels = VAEEncode.vae_encode_crop_pixels(pixels)
-        t = vae.encode_tiled(pixels[:,:,:,:3])
+        t = vae.encode_tiled(pixels[:,:,:,:3], tile_x=tile_size, tile_y=tile_size, )
         return ({"samples":t}, )
 
 class VAEEncodeForInpaint:
@@ -444,7 +474,7 @@ class CheckpointLoaderSimple:
     def load_checkpoint(self, ckpt_name, output_vae=True, output_clip=True):
         ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
         out = comfy.sd.load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, embedding_directory=folder_paths.get_folder_paths("embeddings"))
-        return out
+        return out[:3]
 
 class DiffusersLoader:
     @classmethod
@@ -470,7 +500,7 @@ class DiffusersLoader:
                     model_path = path
                     break
 
-        return comfy.diffusers_load.load_diffusers(model_path, fp16=comfy.model_management.should_use_fp16(), output_vae=output_vae, output_clip=output_clip, embedding_directory=folder_paths.get_folder_paths("embeddings"))
+        return comfy.diffusers_load.load_diffusers(model_path, output_vae=output_vae, output_clip=output_clip, embedding_directory=folder_paths.get_folder_paths("embeddings"))
 
 
 class unCLIPCheckpointLoader:
@@ -569,7 +599,7 @@ class ControlNetLoader:
 
     def load_controlnet(self, control_net_name):
         controlnet_path = folder_paths.get_full_path("controlnet", control_net_name)
-        controlnet = comfy.sd.load_controlnet(controlnet_path)
+        controlnet = comfy.controlnet.load_controlnet(controlnet_path)
         return (controlnet,)
 
 class DiffControlNetLoader:
@@ -585,7 +615,7 @@ class DiffControlNetLoader:
 
     def load_controlnet(self, model, control_net_name):
         controlnet_path = folder_paths.get_full_path("controlnet", control_net_name)
-        controlnet = comfy.sd.load_controlnet(controlnet_path, model)
+        controlnet = comfy.controlnet.load_controlnet(controlnet_path, model)
         return (controlnet,)
 
 
@@ -1578,6 +1608,7 @@ NODE_CLASS_MAPPINGS = {
     "ConditioningCombine": ConditioningCombine,
     "ConditioningConcat": ConditioningConcat,
     "ConditioningSetArea": ConditioningSetArea,
+    "ConditioningSetAreaPercentage": ConditioningSetAreaPercentage,
     "ConditioningSetMask": ConditioningSetMask,
     "KSamplerAdvanced": KSamplerAdvanced,
     "SetLatentNoiseMask": SetLatentNoiseMask,
@@ -1639,6 +1670,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "ConditioningAverage ": "Conditioning (Average)",
     "ConditioningConcat": "Conditioning (Concat)",
     "ConditioningSetArea": "Conditioning (Set Area)",
+    "ConditioningSetAreaPercentage": "Conditioning (Set Area with Percentage)",
     "ConditioningSetMask": "Conditioning (Set Mask)",
     "ControlNetApply": "Apply ControlNet",
     "ControlNetApplyAdvanced": "Apply ControlNet (Advanced)",
diff --git a/server.py b/server.py
index d1295342..57d5a65d 100644
--- a/server.py
+++ b/server.py
@@ -1,6 +1,8 @@
 import os
 import sys
 import asyncio
+import traceback
+
 import nodes
 import folder_paths
 import execution
@@ -10,6 +12,7 @@ import json
 import glob
 import struct
 from PIL import Image, ImageOps
+from PIL.PngImagePlugin import PngInfo
 from io import BytesIO
 
 try:
@@ -79,7 +82,7 @@ class PromptServer():
         if args.enable_cors_header:
             middlewares.append(create_cors_middleware(args.enable_cors_header))
 
-        self.app = web.Application(client_max_size=20971520, middlewares=middlewares)
+        self.app = web.Application(client_max_size=104857600, middlewares=middlewares)
         self.sockets = dict()
         self.web_root = os.path.join(os.path.dirname(
             os.path.realpath(__file__)), "web")
@@ -88,6 +91,8 @@ class PromptServer():
         self.last_node_id = None
         self.client_id = None
 
+        self.on_prompt_handlers = []
+
         @routes.get('/ws')
         async def websocket_handler(request):
             ws = web.WebSocketResponse()
@@ -122,7 +127,7 @@ class PromptServer():
         @routes.get("/embeddings")
         def get_embeddings(self):
             embeddings = folder_paths.get_filename_list("embeddings")
-            return web.json_response(list(map(lambda a: os.path.splitext(a)[0].lower(), embeddings)))
+            return web.json_response(list(map(lambda a: os.path.splitext(a)[0], embeddings)))
 
         @routes.get("/extensions")
         async def get_extensions(request):
@@ -229,13 +234,17 @@ class PromptServer():
 
                 if os.path.isfile(file):
                     with Image.open(file) as original_pil:
+                        metadata = PngInfo()
+                        if hasattr(original_pil,'text'):
+                            for key in original_pil.text:
+                                metadata.add_text(key, original_pil.text[key])
                         original_pil = original_pil.convert('RGBA')
                         mask_pil = Image.open(image.file).convert('RGBA')
 
                         # alpha copy
                         new_alpha = mask_pil.getchannel('A')
                         original_pil.putalpha(new_alpha)
-                        original_pil.save(filepath, compress_level=4)
+                        original_pil.save(filepath, compress_level=4, pnginfo=metadata)
 
             return image_upload(post, image_save_function)
 
@@ -438,6 +447,7 @@ class PromptServer():
             resp_code = 200
             out_string = ""
             json_data =  await request.json()
+            json_data = self.trigger_on_prompt(json_data)
 
             if "number" in json_data:
                 number = float(json_data['number'])
@@ -606,3 +616,15 @@ class PromptServer():
         if call_on_start is not None:
             call_on_start(address, port)
 
+    def add_on_prompt_handler(self, handler):
+        self.on_prompt_handlers.append(handler)
+
+    def trigger_on_prompt(self, json_data):
+        for handler in self.on_prompt_handlers:
+            try:
+                json_data = handler(json_data)
+            except Exception as e:
+                print(f"[ERROR] An error occurred during the on_prompt_handler processing")
+                traceback.print_exc()
+
+        return json_data
diff --git a/web/extensions/core/groupOptions.js b/web/extensions/core/groupOptions.js
new file mode 100644
index 00000000..1d935e90
--- /dev/null
+++ b/web/extensions/core/groupOptions.js
@@ -0,0 +1,167 @@
+import {app} from "../../scripts/app.js";
+
+function setNodeMode(node, mode) {
+    node.mode = mode;
+    node.graph.change();
+}
+
+app.registerExtension({
+    name: "Comfy.GroupOptions",
+    setup() {
+        const orig = LGraphCanvas.prototype.getCanvasMenuOptions;
+        // graph_mouse
+        LGraphCanvas.prototype.getCanvasMenuOptions = function () {
+            const options = orig.apply(this, arguments);
+            const group = this.graph.getGroupOnPos(this.graph_mouse[0], this.graph_mouse[1]);
+            if (!group) {
+                return options;
+            }
+
+            // Group nodes aren't recomputed until the group is moved, this ensures the nodes are up-to-date
+            group.recomputeInsideNodes();
+            const nodesInGroup = group._nodes;
+
+            // No nodes in group, return default options
+            if (nodesInGroup.length === 0) {
+                return options;
+            } else {
+                // Add a separator between the default options and the group options
+                options.push(null);
+            }
+
+            // Check if all nodes are the same mode
+            let allNodesAreSameMode = true;
+            for (let i = 1; i < nodesInGroup.length; i++) {
+                if (nodesInGroup[i].mode !== nodesInGroup[0].mode) {
+                    allNodesAreSameMode = false;
+                    break;
+                }
+            }
+
+            // Modes
+            // 0: Always
+            // 1: On Event
+            // 2: Never
+            // 3: On Trigger
+            // 4: Bypass
+            // If all nodes are the same mode, add a menu option to change the mode
+            if (allNodesAreSameMode) {
+                const mode = nodesInGroup[0].mode;
+                switch (mode) {
+                    case 0:
+                        // All nodes are always, option to disable, and bypass
+                        options.push({
+                            content: "Set Group Nodes to Never",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 2);
+                                }
+                            }
+                        });
+                        options.push({
+                            content: "Bypass Group Nodes",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 4);
+                                }
+                            }
+                        });
+                        break;
+                    case 2:
+                        // All nodes are never, option to enable, and bypass
+                        options.push({
+                            content: "Set Group Nodes to Always",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 0);
+                                }
+                            }
+                        });
+                        options.push({
+                            content: "Bypass Group Nodes",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 4);
+                                }
+                            }
+                        });
+                        break;
+                    case 4:
+                        // All nodes are bypass, option to enable, and disable
+                        options.push({
+                            content: "Set Group Nodes to Always",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 0);
+                                }
+                            }
+                        });
+                        options.push({
+                            content: "Set Group Nodes to Never",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 2);
+                                }
+                            }
+                        });
+                        break;
+                    default:
+                        // All nodes are On Trigger or On Event(Or other?), option to disable, set to always, or bypass
+                        options.push({
+                            content: "Set Group Nodes to Always",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 0);
+                                }
+                            }
+                        });
+                        options.push({
+                            content: "Set Group Nodes to Never",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 2);
+                                }
+                            }
+                        });
+                        options.push({
+                            content: "Bypass Group Nodes",
+                            callback: () => {
+                                for (const node of nodesInGroup) {
+                                    setNodeMode(node, 4);
+                                }
+                            }
+                        });
+                        break;
+                }
+            } else {
+                // Nodes are not all the same mode, add a menu option to change the mode to always, never, or bypass
+                options.push({
+                    content: "Set Group Nodes to Always",
+                    callback: () => {
+                        for (const node of nodesInGroup) {
+                            setNodeMode(node, 0);
+                        }
+                    }
+                });
+                options.push({
+                    content: "Set Group Nodes to Never",
+                    callback: () => {
+                        for (const node of nodesInGroup) {
+                            setNodeMode(node, 2);
+                        }
+                    }
+                });
+                options.push({
+                    content: "Bypass Group Nodes",
+                    callback: () => {
+                        for (const node of nodesInGroup) {
+                            setNodeMode(node, 4);
+                        }
+                    }
+                });
+            }
+
+            return options
+        }
+    }
+});
diff --git a/web/extensions/core/linkRenderMode.js b/web/extensions/core/linkRenderMode.js
index 1e9091ec..fb4df423 100644
--- a/web/extensions/core/linkRenderMode.js
+++ b/web/extensions/core/linkRenderMode.js
@@ -9,7 +9,7 @@ const ext = {
 			name: "Link Render Mode",
 			defaultValue: 2,
 			type: "combo",
-			options: LiteGraph.LINK_RENDER_MODES.map((m, i) => ({
+			options: [...LiteGraph.LINK_RENDER_MODES, "Hidden"].map((m, i) => ({
 				value: i,
 				text: m,
 				selected: i == app.canvas.links_render_mode,
diff --git a/web/extensions/core/widgetInputs.js b/web/extensions/core/widgetInputs.js
index d9eaf8a0..f9a5b727 100644
--- a/web/extensions/core/widgetInputs.js
+++ b/web/extensions/core/widgetInputs.js
@@ -5,7 +5,7 @@ const CONVERTED_TYPE = "converted-widget";
 const VALID_TYPES = ["STRING", "combo", "number", "BOOLEAN"];
 
 function isConvertableWidget(widget, config) {
-	return VALID_TYPES.includes(widget.type) || VALID_TYPES.includes(config[0]);
+	return (VALID_TYPES.includes(widget.type) || VALID_TYPES.includes(config[0])) && !widget.options?.forceInput;
 }
 
 function hideWidget(node, widget, suffix = "") {
@@ -16,8 +16,12 @@ function hideWidget(node, widget, suffix = "") {
 	widget.type = CONVERTED_TYPE + suffix;
 	widget.serializeValue = () => {
 		// Prevent serializing the widget if we have no input linked
-		const { link } = node.inputs.find((i) => i.widget?.name === widget.name);
-		if (link == null) {
+		if (!node.inputs) {
+			return undefined;
+		}
+		let node_input = node.inputs.find((i) => i.widget?.name === widget.name);
+
+		if (!node_input || !node_input.link) {
 			return undefined;
 		}
 		return widget.origSerializeValue ? widget.origSerializeValue() : widget.value;
@@ -103,6 +107,9 @@ app.registerExtension({
 				let toInput = [];
 				let toWidget = [];
 				for (const w of this.widgets) {
+					if (w.options?.forceInput) {
+						continue;
+					}
 					if (w.type === CONVERTED_TYPE) {
 						toWidget.push({
 							content: `Convert ${w.name} to widget`,
@@ -130,6 +137,20 @@ app.registerExtension({
 			return r;
 		};
 
+		const origOnNodeCreated = nodeType.prototype.onNodeCreated
+		nodeType.prototype.onNodeCreated = function () {
+			const r = origOnNodeCreated ? origOnNodeCreated.apply(this) : undefined;
+			if (this.widgets) {
+				for (const w of this.widgets) {
+					if (w?.options?.forceInput) {
+						const config = nodeData?.input?.required[w.name] || nodeData?.input?.optional?.[w.name] || [w.type, w.options || {}];
+						convertToInput(this, w, config);
+					}
+				}
+			}
+			return r;
+		}
+
 		// On initial configure of nodes hide all converted widgets
 		const origOnConfigure = nodeType.prototype.onConfigure;
 		nodeType.prototype.onConfigure = function () {
@@ -137,7 +158,7 @@ app.registerExtension({
 
 			if (this.inputs) {
 				for (const input of this.inputs) {
-					if (input.widget) {
+					if (input.widget && !input.widget.config[1]?.forceInput) {
 						const w = this.widgets.find((w) => w.name === input.widget.name);
 						if (w) {
 							hideWidget(this, w);
@@ -374,7 +395,7 @@ app.registerExtension({
 				}
 
 				for (const k in config1[1]) {
-					if (k !== "default") {
+					if (k !== "default" && k !== 'forceInput') {
 						if (config1[1][k] !== config2[1][k]) {
 							return false;
 						}
diff --git a/web/lib/litegraph.core.js b/web/lib/litegraph.core.js
index 356c71ac..4a21a1b3 100644
--- a/web/lib/litegraph.core.js
+++ b/web/lib/litegraph.core.js
@@ -6233,11 +6233,17 @@ LGraphNode.prototype.executeAction = function(action)
 																					,posAdd:[!mClikSlot_isOut?-30:30, -alphaPosY*130] //-alphaPosY*30]
 																					,posSizeFix:[!mClikSlot_isOut?-1:0, 0] //-alphaPosY*2*/
 																				});
-								
+							skip_action = true;
 						}
 					}
 				}
 			}
+
+			if (!skip_action && this.allow_dragcanvas) {
+            	//console.log("pointerevents: dragging_canvas start from middle button");
+            	this.dragging_canvas = true;
+            }
+
         	
         } else if (e.which == 3 || this.pointer_is_double) {
 			
@@ -11523,7 +11529,7 @@ LGraphNode.prototype.executeAction = function(action)
                     if (timeout) {
                         clearInterval(timeout);
                     }
-                    timeout = setTimeout(refreshHelper, 250);
+                    timeout = setTimeout(refreshHelper, 10);
                     return;
                 }
                 e.preventDefault();
diff --git a/web/scripts/app.js b/web/scripts/app.js
index 6a2c6329..a3661da6 100644
--- a/web/scripts/app.js
+++ b/web/scripts/app.js
@@ -667,11 +667,40 @@ export class ComfyApp {
 	}
 
 	/**
-	 * Adds a handler on paste that extracts and loads workflows from pasted JSON data
+	 * Adds a handler on paste that extracts and loads images or workflows from pasted JSON data
 	 */
 	#addPasteHandler() {
 		document.addEventListener("paste", (e) => {
-			let data = (e.clipboardData || window.clipboardData).getData("text/plain");
+			let data = (e.clipboardData || window.clipboardData);
+			const items = data.items;
+
+			// Look for image paste data
+			for (const item of items) {
+				if (item.type.startsWith('image/')) {
+					var imageNode = null;
+
+					// If an image node is selected, paste into it
+					if (this.canvas.current_node &&
+						this.canvas.current_node.is_selected &&
+						ComfyApp.isImageNode(this.canvas.current_node)) {
+						imageNode = this.canvas.current_node;
+					}
+
+					// No image node selected: add a new one
+					if (!imageNode) {
+						const newNode = LiteGraph.createNode("LoadImage");
+						newNode.pos = [...this.canvas.graph_mouse];
+						imageNode = this.graph.add(newNode);
+						this.graph.change();
+					}
+					const blob = item.getAsFile();
+					imageNode.pasteFile(blob);
+					return;
+				}
+			}
+
+			// No image found. Look for node data
+			data = data.getData("text/plain");
 			let workflow;
 			try {
 				data = data.slice(data.indexOf("{"));
@@ -687,9 +716,33 @@ export class ComfyApp {
 			if (workflow && workflow.version && workflow.nodes && workflow.extra) {
 				this.loadGraphData(workflow);
 			}
+			else {
+				if (e.target.type === "text" || e.target.type === "textarea") {
+					return;
+				}
+
+				// Litegraph default paste
+				this.canvas.pasteFromClipboard();
+			}
+
+
+		});
+	}
+
+
+	/**
+	 * Adds a handler on copy that serializes selected nodes to JSON
+	 */
+	#addCopyHandler() {
+		document.addEventListener("copy", (e) => {
+			// copy
+			if (this.canvas.selected_nodes) {
+			    this.canvas.copyToClipboard();
+			}
 		});
 	}
 
+
 	/**
 	 * Handle mouse
 	 *
@@ -745,12 +798,6 @@ export class ComfyApp {
 		const self = this;
 		const origProcessKey = LGraphCanvas.prototype.processKey;
 		LGraphCanvas.prototype.processKey = function(e) {
-			const res = origProcessKey.apply(this, arguments);
-
-			if (res === false) {
-				return res;
-			}
-
 			if (!this.graph) {
 				return;
 			}
@@ -761,9 +808,10 @@ export class ComfyApp {
 				return;
 			}
 
-			if (e.type == "keydown") {
+			if (e.type == "keydown" && !e.repeat) {
+
 				// Ctrl + M mute/unmute
-				if (e.keyCode == 77 && e.ctrlKey) {
+				if (e.key === 'm' && e.ctrlKey) {
 					if (this.selected_nodes) {
 						for (var i in this.selected_nodes) {
 							if (this.selected_nodes[i].mode === 2) { // never
@@ -776,7 +824,8 @@ export class ComfyApp {
 					block_default = true;
 				}
 
-				if (e.keyCode == 66 && e.ctrlKey) {
+				// Ctrl + B bypass
+				if (e.key === 'b' && e.ctrlKey) {
 					if (this.selected_nodes) {
 						for (var i in this.selected_nodes) {
 							if (this.selected_nodes[i].mode === 4) { // never
@@ -788,6 +837,28 @@ export class ComfyApp {
 					}
 					block_default = true;
 				}
+
+				// Ctrl+C Copy
+				if ((e.key === 'c') && (e.metaKey || e.ctrlKey)) {
+					if (e.shiftKey) {
+						this.copyToClipboard(true);
+						block_default = true;
+					}
+					// Trigger default onCopy
+					return true;
+				}
+
+				// Ctrl+V Paste
+				if ((e.key === 'v') && (e.metaKey || e.ctrlKey)) {
+					if (e.shiftKey) {
+						this.pasteFromClipboard(true);
+						block_default = true;
+					}
+					else {
+						// Trigger default onPaste
+						return true;
+					}
+				}
 			}
 
 			this.graph.change();
@@ -798,7 +869,8 @@ export class ComfyApp {
 				return false;
 			}
 
-			return res;
+			// Fall through to Litegraph defaults
+			return origProcessKey.apply(this, arguments);
 		};
 	}
 
@@ -994,6 +1066,10 @@ export class ComfyApp {
 		api.addEventListener("execution_start", ({ detail }) => {
 			this.runningNodeId = null;
 			this.lastExecutionError = null
+			this.graph._nodes.forEach((node) => {
+				if (node.onExecutionStart)
+					node.onExecutionStart()
+			})
 		});
 
 		api.addEventListener("execution_error", ({ detail }) => {
@@ -1110,6 +1186,7 @@ export class ComfyApp {
 		this.#addDrawGroupsHandler();
 		this.#addApiUpdateHandlers();
 		this.#addDropHandler();
+		this.#addCopyHandler();
 		this.#addPasteHandler();
 		this.#addKeyboardHandler();
 
@@ -1151,22 +1228,25 @@ export class ComfyApp {
 						const inputData = inputs[inputName];
 						const type = inputData[0];
 
-						if(inputData[1]?.forceInput) {
-							this.addInput(inputName, type);
+						let widgetCreated = true;
+						if (Array.isArray(type)) {
+							// Enums
+							Object.assign(config, widgets.COMBO(this, inputName, inputData, app) || {});
+						} else if (`${type}:${inputName}` in widgets) {
+							// Support custom widgets by Type:Name
+							Object.assign(config, widgets[`${type}:${inputName}`](this, inputName, inputData, app) || {});
+						} else if (type in widgets) {
+							// Standard type widgets
+							Object.assign(config, widgets[type](this, inputName, inputData, app) || {});
 						} else {
-							if (Array.isArray(type)) {
-								// Enums
-								Object.assign(config, widgets.COMBO(this, inputName, inputData, app) || {});
-							} else if (`${type}:${inputName}` in widgets) {
-								// Support custom widgets by Type:Name
-								Object.assign(config, widgets[`${type}:${inputName}`](this, inputName, inputData, app) || {});
-							} else if (type in widgets) {
-								// Standard type widgets
-								Object.assign(config, widgets[type](this, inputName, inputData, app) || {});
-							} else {
-								// Node connection inputs
-								this.addInput(inputName, type);
-							}
+							// Node connection inputs
+							this.addInput(inputName, type);
+							widgetCreated = false;
+						}
+
+						if(widgetCreated && inputData[1]?.forceInput && config?.widget) {
+							if (!config.widget.options) config.widget.options = {};
+							config.widget.options.forceInput = inputData[1].forceInput;
 						}
 					}
 
diff --git a/web/scripts/ui.js b/web/scripts/ui.js
index 86e2a1c4..f39939bf 100644
--- a/web/scripts/ui.js
+++ b/web/scripts/ui.js
@@ -431,10 +431,12 @@ class ComfySettingsDialog extends ComfyDialog {
 class ComfyList {
 	#type;
 	#text;
+	#reverse;
 
-	constructor(text, type) {
+	constructor(text, type, reverse) {
 		this.#text = text;
 		this.#type = type || text.toLowerCase();
+		this.#reverse = reverse || false;
 		this.element = $el("div.comfy-list");
 		this.element.style.display = "none";
 	}
@@ -451,7 +453,7 @@ class ComfyList {
 					textContent: section,
 				}),
 				$el("div.comfy-list-items", [
-					...items[section].map((item) => {
+					...(this.#reverse ? items[section].reverse() : items[section]).map((item) => {
 						// Allow items to specify a custom remove action (e.g. for interrupt current prompt)
 						const removeAction = item.remove || {
 							name: "Delete",
@@ -529,7 +531,7 @@ export class ComfyUI {
 		this.batchCount = 1;
 		this.lastQueueSize = 0;
 		this.queue = new ComfyList("Queue");
-		this.history = new ComfyList("History");
+		this.history = new ComfyList("History", "history", true);
 
 		api.addEventListener("status", () => {
 			this.queue.update();
@@ -617,7 +619,9 @@ export class ComfyUI {
 				]),
 			]),
 			$el("div", {id: "extraOptions", style: {width: "100%", display: "none"}}, [
-				$el("label", {innerHTML: "Batch count"}, [
+				$el("div",[
+
+					$el("label", {innerHTML: "Batch count"}),
 					$el("input", {
 						id: "batchCountInputNumber",
 						type: "number",
@@ -639,14 +643,23 @@ export class ComfyUI {
 							this.batchCount = i.srcElement.value;
 							document.getElementById("batchCountInputNumber").value = i.srcElement.value;
 						},
+					}),		
+				]),
+
+				$el("div",[
+					$el("label",{
+						for:"autoQueueCheckbox",
+						innerHTML: "Auto Queue"
+						// textContent: "Auto Queue"
 					}),
 					$el("input", {
 						id: "autoQueueCheckbox",
 						type: "checkbox",
 						checked: false,
-						title: "automatically queue prompt when the queue size hits 0",
+						title: "Automatically queue prompt when the queue size hits 0",
+						
 					}),
-				]),
+				])
 			]),
 			$el("div.comfy-menu-btns", [
 				$el("button", {
diff --git a/web/scripts/widgets.js b/web/scripts/widgets.js
index adf5f26f..45ac9b89 100644
--- a/web/scripts/widgets.js
+++ b/web/scripts/widgets.js
@@ -76,7 +76,7 @@ export function addValueControlWidget(node, targetWidget, defaultValue = "random
 				targetWidget.value = max;
 		}
 	}
-	return valueControl;	
+	return valueControl;
 };
 
 function seedWidget(node, inputName, inputData, app) {
@@ -299,11 +299,17 @@ export const ComfyWidgets = {
 		const defaultVal = inputData[1].default || "";
 		const multiline = !!inputData[1].multiline;
 
+		let res;
 		if (multiline) {
-			return addMultilineWidget(node, inputName, { defaultVal, ...inputData[1] }, app);
+			res = addMultilineWidget(node, inputName, { defaultVal, ...inputData[1] }, app);
 		} else {
-			return { widget: node.addWidget("text", inputName, defaultVal, () => {}, {}) };
+			res = { widget: node.addWidget("text", inputName, defaultVal, () => {}, {}) };
 		}
+
+		if(inputData[1].dynamicPrompts != undefined)
+			res.widget.dynamicPrompts = inputData[1].dynamicPrompts;
+
+		return res;
 	},
 	COMBO(node, inputName, inputData) {
 		const type = inputData[0];
@@ -381,11 +387,12 @@ export const ComfyWidgets = {
 			}
 		});
 
-		async function uploadFile(file, updateNode) {
+		async function uploadFile(file, updateNode, pasted = false) {
 			try {
 				// Wrap file in formdata so it includes filename
 				const body = new FormData();
 				body.append("image", file);
+				if (pasted) body.append("subfolder", "pasted");
 				const resp = await api.fetchApi("/upload/image", {
 					method: "POST",
 					body,
@@ -393,15 +400,17 @@ export const ComfyWidgets = {
 
 				if (resp.status === 200) {
 					const data = await resp.json();
-					// Add the file as an option and update the widget value
-					if (!imageWidget.options.values.includes(data.name)) {
-						imageWidget.options.values.push(data.name);
+					// Add the file to the dropdown list and update the widget value
+					let path = data.name;
+					if (data.subfolder) path = data.subfolder + "/" + path;
+
+					if (!imageWidget.options.values.includes(path)) {
+						imageWidget.options.values.push(path);
 					}
 
 					if (updateNode) {
-						showImage(data.name);
-
-						imageWidget.value = data.name;
+						showImage(path);
+						imageWidget.value = path;
 					}
 				} else {
 					alert(resp.status + " - " + resp.statusText);
@@ -454,6 +463,16 @@ export const ComfyWidgets = {
 			return handled;
 		};
 
+		node.pasteFile = function(file) {
+			if (file.type.startsWith("image/")) {
+				const is_pasted = (file.name === "image.png") &&
+								  (file.lastModified - Date.now() < 2000);
+				uploadFile(file, true, is_pasted);
+				return true;
+			}
+			return false;
+		}
+
 		return { widget: uploadWidget };
 	},
 };
diff --git a/web/style.css b/web/style.css
index 5b6b9ec5..692fa31d 100644
--- a/web/style.css
+++ b/web/style.css
@@ -88,7 +88,7 @@ body {
 	top: 50%;
 	right: 0;
 	text-align: center;
-	z-index: 100;
+	z-index: 999;
 	width: 170px;
 	display: flex;
 	flex-direction: column;