Once ComfyUI is installed and running, adding workflows is as easy as dragging and dropping images or workflows created by ComfyUI into the empty area of the browser window.
To follow along with this tutorial, download [the workflow](https://raw.githubusercontent.com/wyrde/wyrde-comfyui-workflows/main/basics/building-up/basic-workflow-v03.json) and save it. Then use the `[ Load ]` button in ComfyUI to load it.
* The _Seed_ is a number used for randomness. All things being equal, a workflow with the same seed will spit out the same image.
* If the seed doesn't change and nothing else changes, ComfyUI won't even process the workflow. The result is the same as last time. If nothing else changes and the seed does, then new images will be generated. They can vary wildly from one to the next.
* This makes it easy to tell if nothing has changed.
* The line under the seed is how new seeds are generated.
* fixed: they aren't. Seed stays the same. Useful for checking othetr parameters and generating the same basic image.
* Increment/Decriment: the seed changes by +/- 1.
* Random: Holdy Changing Numbers, Batman!
* values are changed by clicking the arrows to scroll through the options or clicking directly on a value to see a list.
Click "Queue Prompt" in the box on the side of the window to generate an image. If the same settings are used from the workflow above, it'll look remarkably like
* a noodle can be pulled from the _VAE Decode_ node's input and dropped to show a Vae Loader in the menu, but that'll be messy. Let's save messy workflows for crazy vixens like comfy and mike.
Download a VAE from [stabilityai](https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main) and drop it in `ComfyUI\models\vae`. Get the pruned.safetensors file.
* Don't click the filename (it results in a page that is confusing until eyes finally spot the "download" link), click to the right. <imgsrc="./pix/downloadvae.png"width="75%"align="middle">.
In this particular case, there isn't a huge difference in making a new version with the vae linked. For an idea of what differences vae make, check out [this comparison](https://github.com/wyrde/wyrde-comfyui-workflows/tree/main/compare/model-vae). The ones with bad vae are faded and washed out. Good vae are vibrant.
Note
* Automatic1111's method of matching vae file names to models and placing them in the model folder doesn't matter to ComfyUI. Either the VAE is loaded directly or the one built into the model is used.
* [Here's some good info on VAE](https://rentry.org/sdvae)
Lora (and the varients) are cool mini-models that are used to alter a bigger model. Think of them like the trojan horse, but everyone is happy with the result. Usually.
[here's a lora](https://civitai.com/models/44960?modelVersionId=49584) to test with. Drop it in the `ComfyUI\models\lora\` folder.
* as with other files, hit F5 to refresh the file lists
* ignore the text in the lora page for now, what is first seen is for version 3.0 and version 1 is what we're using.
Adding a lora is a bit more tricky than a vae loader, because they go between the model and the Ksampler.
* Unlike automatic1111, the lora aren't put in the prompt.
Make a little room for the _lora Loader_ node.
* Between the _CLIP Text_ nods and _Load Checkpoint_ hold down the control key and drag with the left mouse button. It will draw a box. When Left Mouse is released, everything inside the box is selected.
* Now hold down shift and Left Click one of the selected nodes. All the nodes will be dragged as a group when the mouse is dragged.
* It is better to make room by dragging nodes right than left
* Refreshing (F5) the window zooms to the barely visiable blue box where the beginning nodes are.
* it gets annoying to pan left all the time
Now drop a _Lora Loader_ in the empty spot.
* Connect the _Load Checkpoint_ model and clip outputs to the Load Lora.
* The _Load LoRA_'s model output goes to the model reroute (to the KSampler model input)
* this will autmatically detach it from the Load Checkpoint.
* The _Load LoRA_'s CLIP output is noodled to both the CLIP nodes' inputs.
* this will autmatically detach them from the Load Checkpoint.
* Select _mPixel_v10pixelArt.safetensors_ for the lora_name field.
So, an important thing about stable diffusion is the models are trained on small images. 512x512 pixels is the standard size for most SD1.5 based models (768 for a few). SD2.x models can be either 512px or 768px, depending on the one chosen.
But what does that mean for regular users? Most people want a much larger image (1920x1080 for example). The thing is, just changing the Latent Image size to 1920x1080 tends to go horribly, horribly wrong. That's because stable diffusion doesn't really understand "size" or "composition". When it sees a huge canvase size, it tries to fill _every part_ of it with the prompt.
* the previous prompt and configuration with a 1920x1080 size latent:
But fear not, there are a few techniques to increase image size from 512px to something more grandiose.
* Latent Upscale: this takes a latent image and makes it bigger. The result is okay, but the larger image is missing much of the detail that is possible.
* Pixel space upscale: Tends to look beter than latent upscale, but is still missing details.
* Pixel space upscale with model: There are special models designed for a better upscale result. Some even extrapolate details.
* High Resolution Fixing: Uusually abbreviated to HR-Fix or even furthetr to HRF, this method is a combination of many of the above. It uses many steps, takes longer, but results in high-detail images. There's the added bonuse that stable diffusion can even "fix" broken parts of the image! Sometimes.
* Latent space is the zone of chaos and randomness stable diffusion works in. These aren't really images like human eyes see, but math-magic representations the computer works with.
* pixel space is an actual image, rendered in pixels. The things human eyes and graphic arts programs like to work with.
* VAE Decoding and Encoding is used to translate a latent image to pixel and back.
* Some proccesses can only be done in latent space (like KSampling) and others only in pixel space (color correction, model upscaling)
There are some caveats to HR-fixing.
* work in as many small steps as you can stand.
* stick to multiples of 64px. This can be challenging with images of different aspect ratios. There are some custom nodes that help with this.
* if the aspect ratio makes multiples of 64px challenging, do the image upscales in pixel space
* due to the math involved, even then it is best to do multiples of 8 pixels.
* it is better (and faster) to go through many low-resolution latent space samples and _then_ do a large pixel upscale. This gives lots of detail for the pixel upscale to work with.
But first, let's make a basic HR-Fix.
* at the right side of the workflow, move the VAE Decode and Save Image nodes to the right (try and grab the VAE reroute on the bottom too).
* about as far as the distance they currently occupy, maybe a touch more.
* put a _Upscale Latent_ node next to the Ksampler
* connect the LATENT output on the Ksampler to the Samples input on the _Upscale Latent_ node.
* put a _KSampler_ node between the Upscale Latent and the VAE Decode.
* connect he LATENT output from the _Upscale Latent_ node to the latent_image input on the new _KSampler_ node.
* connect the LATEN output on the ksampler to the samples input on the _VAE Decode_ node.
* this will also erase the noodle from the old ksampler to the vae decode
* On the Upscale Latent node, increase wideth and hight by 64pixels. This is easy by clicking the arrows on the right, next to the numbers.
But wait! We're not done. The new KSampler still has a bunch of inputs that are empty. That's because the sampler needs to know what to do with the latent image.
See the model rerout hanging out at the corner of the positive prompt?
* lect click to activate it.
* Ctrl-C to clone it.
* move mouse a little to the right (above the older ksmampler is fine)
* adjust the denoise down to more closely match the original latent while still adding detail.
* 0.500 is _generally_ a good number for the first "HR Fix"
* click Queue Prompt again.
* Instead of running the whole workflow, comfyui should start at the second KSampler. This is because the there were no changes earlier in the workflow.
* and decrease denoise a little. For this one, .450 is good.
* each ksampler in a HRF will decrease the noise
* though for latent space, about .2 is the lowest. Usually.
[here's the current workflow](basic-wf-vae-lora-latemt-upscale-x2.json)
More latent HRFs will gradually increase the output image while adding details. But let's stop here and add some pixel space HRFs. Onewards, noble steed!
First, we need to translate from latent space to pixel space.
* Select the VAE Decode and its VAE reroute below and clone them over by the last ksampler.
<imgsrc="./pix/pixel space hrf 1.png"width="80%"align="middle">
* drag the output nodes to the right again
<imgsrc="./pix/pixel space hrf 2.png"width="80%"align="middle">
* Now add a _Upscale Image_ node
<imgsrc="./pix/pixel space hrf 3.png"width="80%"align="middle">
<imgsrc="./pix/pixel space hrf 4.png"width="80%"align="middle">
<imgsrc="./pix/pixel space hrf 5.png"width="80%"align="middle">
* After the upscale node, we need a _VAE Encode_ node.
<imgsrc="./pix/pixel space hrf 8.png"width="80%"align="middle">
<imgsrc="./pix/pixel space hrf 9.png"width="80%"align="middle">
* Select the reroute nodes and ksampler and clone them into the space after the new encode node
<imgsrc="./pix/pixel space hrf 10.png"width="80%"align="middle">
* Connect all the noodles
<imgsrc="./pix/pixel space hrf 11.png"width="80%"align="middle">
<imgsrc="./pix/pixel space hrf 12.png"width="80%"align="middle">
<imgsrc="./pix/pixel space hrf 13.png"width="80%"align="middle">
* During this process, the connection to the first VAE reroute all the way on the left will likely be lost. Don't worry! Pan over to the left by holding down space while moving the mouse.
* space + mouse = pan.
* release space to move mouse back
* space + mouse to pan again
* start dragging the noodle from the leftmost vae reroute
* space and mouse to keep going
* and attach it to the newer VAE reroute.
* the workflow should now look something like this:
* unlike latent space HR Fixes, after a pixel upscale the denoise doesn't need decreased _as much_. In this case, leaving it at .5 is okay.
* generally speaking, sampling after pixel space upscales doesn't need decreased below 0.400. Generally. There's always exceptions. That's the norm with Stable Diffusion.
Improving images with HiRez Fixes is one thing, but what about simply making it bigger? That's where upscaling comes in. As mentioned before, jumping from 512px to 1080p and higher isn't advised due to the lack of detail. Each HR Fix also gives stable diffues the chance to correct mistakes. (It doesn't always, though. Stable Diffusion is a contrary baast.)
Once an image has some detail, one of the best ways to upscale it further is with an _Upscale Model_. There's a whole bunch of them [here](https://upscale.wiki/wiki/Model_Database), but we'll keep it easy with PSNRx2. Follow [https://huggingface.co/wyrde/upscales/tree/main/apache2](https://huggingface.co/wyrde/upscales/tree/main/apache2) and click the LFS button to download.
Otherwise known as Textual Inversions, embeddings are a bit different than the usual additions to a comfy workflow. Instead of a node, they are added directly to a prompt. This is due to the nature of an embedding, they are specialized models which provide a specific meaning to a token.
Wait, tokens?
While prompts are human readable, they are not read in a human fashion by stable diffusion. They are translated into _tokens_, and stable diffusion assigns meaning to them according to associations in the model (including lora, which come before the prompts).
Tokens are placed in an prompt with the keyword `embedding:` followed by the embedding's filename. For example, a popular _negative_ embedding is EasyNegative.pt.
` embedding:EasyNegative`
There is no space between `embedding:` and `EasyNegative`, it is all one "word." The filename extension (`.pt` in this case) can be left off if desired.
Embeddings can be placed in positive or negative prompts, as desired. Much like lora, trigger words aren't strictly neceessary, but can add additional "weight" to the prompt.
