From d04e5e176cc7cdbb9689a5043fd356d6070788d3 Mon Sep 17 00:00:00 2001 From: wyrde Date: Thu, 6 Apr 2023 16:44:14 -0400 Subject: [PATCH] Update README.md hirez fix totorial --- basics/was-nodes-start/README.md | 35 +++++++++++++++++++++++++++++--- 1 file changed, 32 insertions(+), 3 deletions(-) diff --git a/basics/was-nodes-start/README.md b/basics/was-nodes-start/README.md index dc3562f..72dcdcd 100644 --- a/basics/was-nodes-start/README.md +++ b/basics/was-nodes-start/README.md @@ -28,12 +28,11 @@ Back to img2img. * The _node's_ "mode" can be changed from "resize" to "rescale" to easily reduce larger images as well. - ## Prompts The workflow includes some basic propmts. Prompts are split into 2 types, positive and negative. * Positive prompts are what what you hope to see in the result. - - (Positibe prompts are often referred to simply as _prompts_, but sometimes _prompts_ means both positive and negative prompts. Confusing? Welcome to Stable Diffusion!) -* Negative prompts are items which the user hopes **don't** show up in the result. + - (Positive prompts are often referred to simply as _prompts_, but sometimes _prompts_ means both positive and negative prompts. Confusing? Welcome to Stable Diffusion!) +* Negative prompts are items the user hopes **don't** show up in the result. * There are all sorts of guides on making good prompts. - Basically, a good prompt has a subject and conditionals. - Subject is the main thing desired in the result, such as `adorable cat girl wearing a striped bikini` @@ -41,6 +40,36 @@ The workflow includes some basic propmts. Prompts are split into 2 types, positi - `absurdres, vibrant, photograph, photographic photo, photogenic, rule of thirds, anime, illustration, medium shot, Azulejo, finely detailed, realistic, masterpiece, best quality, illustration, soft focus, HDR 8K, artstation, pixv, unreal engine 5, SFW, ` is a bunch of information about the type of image desired. + as an aside, _Azulejo_ is a Portuguese tile which adds an interesting style to images with many models. - Negative prompts are much shorter. Usually. Some real doozies exist out there, generally fighting specific things showing up in results. For the basics, `bad hands, NSFW, nude` does well. A couple of those can be removed for more risqué images. (: + + Stable Diffusion is often a contrary beast and seems to fight tooth and nail against negativet prompting. + +## Click on _Queue Prompt_ +The most important pat is, of course, making the images. Click on _Queue Prompt_ to start the process. If ComfyUI doesn't like how nodes are noodles, a bright red error will appear. +If it does, various nodes will glow green as it goes through the process. Eventually a picture will appear way at the end. + +## Extras +What's with the _Latent Upscale by Factor (WAS)_ and second _Image Resize_ at the end? They're there to make it easy to hook in HiRez fixes. Also, I felt like it. At the moment they're not doing much. Changing the _factors_ can easily increase the final image size, but it won't look great. + +The basic of Hi-Rez Fixing: +* Move the image saving nodes further to the right. _Shift click_ or _ctrl+drag_ to select multiple nodes. +* Move them a bit further. Maybe a litle more. +* Space bar can be used to pan, even when nodes are "grabbed". +* Holding down _shift_ will snap the selected nodes to the grid. + - Grid size can be configured in the cog wheel on the main Queue Prompt box. +* Now click on empty space to _deselect_ the nodes. +* Return to the _KSampler (WAS)_ node +* If the _Latent Upscale_ node was left next to it, great! If not, move it back. +* Change the factor to 2. +* Now select the _KSampler_ node. Ctrl+C to copy it. +* Click empty space to the right of the _Latent Uspcale_ node. Ctrl+V to paste in the sampler. + - change _denoise_ to around `0.400`. (The decimal is Very Important™) + - for a very basic hi-rez fix, the other settings are okay. +* Connect the LATENT dot on the _Latent Upscale to the Latent_image dot on the new _KSampler_. Connect the other noodles from the various objects to the left of the original _KSampler_. + - bonus points using more redirects to manage the noodles. Ctrl+C and V to copy paste redirects. + - The redirect grab spot is a little buggy. Aim for the top half. +* Connect the LATENT dot on the new _KSampler_ to the samples dot on the _VAEDecode_ node (the old noodle vanishes if it wasn't disconnected earlier). +* Click Queue Prompt to "fix" the previous image. If comfyUI wasn't restarted at some point, it'll use the old data and start on the new _KSampler_. Otherwise it'll start over from the beginning. +* The resulting image is larger, will look a little different (it was sampled), and be of decent quality. + ## resources