How to install/use the Cascade finetune Prototype For Smoothbrains
(And use the custom gradio GUI to avoid seeing nodes)
An Unofficial Guide ::
Resonance Cascade Info Rentry - https://rentry.org/resonance_cascade (by TheOneAndOnlyGod from FG discord)
Lora Training how-to - https://rentry.org/resonance-lora-training
Remember to donate if you like the prototype, donations will give us the final product bake much faster!
SKIP TO 'HOW TO INSTALL' IF YOU PLAN TO USE THE CAPGUI
-- You can use normal ComfyUI without the CAPGUI, but this guide isn't focused on that. For those who wish to, you can get the Resonance prototype TE & Stage C models here:
https://rentry.org/resonancemodel
https://huggingface.co/stabilityai/stable-cascade/tree/main
NEW: CAPGUI now has drag and drop image workflows for Comfyui, so you can use it as an alternative to get workflows for base comfy
NEW: A three pass hi-res fix workflow by Anon on the FG discord for those who want higher quality images: https://files.catbox.moe/2fy9im.json
NEW: Replace Stage_A with 'stage-a-ft-hq' in the workflow for higher quality images; reduced pixilation, better lines and increased background quality. Note - The CAPGUI will now install this option during set up
https://huggingface.co/madebyollin/stage-a-ft-hq/blob/main/stage_a_ft_hq.safetensors
My basic attempt at a simple ComfyUI workflow .json: https://files.catbox.moe/xsa4vy.json
Updated: Weight-Respecting ComfyUI workflow .json: https://files.catbox.moe/td65z8.json
Note: To properly implement prompt weights in ComfyUi, you will need to install the custom node "ComfyUI_ADV_CLIP_emb" https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb
The weight-respecting workflow uses "ComfyUI Prompt Control" and recommends prepending prompts with (STYLE comfy++) as it may not function as intended with others.
https://github.com/asagi4/comfyui-prompt-control
- Cleaner Ksampler output settings, from Anon on /hdg/ :
Maybe just start off with changing your sampler settings. dpmpp_sde works well at around 12 steps at base res, cfg 5-7, simple scheduler. For Stage B dpmpp_sde also works at 3 steps, cfg 1.1-1.5 and normal scheduler. cfg for stage B mostly affects fine details like eyes
https://files.catbox.moe/nqhcv0.png
Before/After Sampler Settings Change [NSFW] - https://files.catbox.moe/3adwjx.png
YOU DO NOT NEED TO DOWNLOAD ANY OF THE ABOVE IF YOU ARE USING THE CAP GUI EASY INSTALLER, IT WILL DOWNLOAD EVERYTHING FOR YOU!
*the CAPGUI now supports prompt control, improved inpainting, i2i and more!
How to Install
If you'd like images to help guide you through the process, check out TheOneAndOnlyGod's https://rentry.org/resonance_cascade CAPGUI section
Step 1:
- Install ComfyUI. If you already have it you can either make a separate install or use your existing one. Forge and A1111 do not currently support Cascade, so using Comfy is necessary.
https://github.com/comfyanonymous/ComfyUI/- Make sure ComfyUI is up to date by running the update_comfyui batch file in the 'update' folder
- Now, run ComfyUI by using the run_ batch file. It should open in your browser with the 'URL' displaying as IP address:PORT numbers by default
Step 2:
- Git pull Jordach's CAPGUI, this is what will prevent you from having to use those super scary nodes
https://github.com/Jordach/CAPGUI
Step 3:
- Run the "start_comfy_installer" windows batch file that is included with the CAPGUI
Step 4:
- CAPGUI Easy Installer should open in your browser window. Fill it in:
ComfyUI Path:
The location on your computer that your ComfyUI is installed to. For example, mine is:C:\Users\B\Downloads\new_ComfyUI_windows_portable_nvidia_cu121_or_cpu\ComfyUI_windows_portable\ComfyUI
ComfyUI IP Address & Port:
Leave at the default setting unless you have changed them. They should be the same as what the 'URL' shows in your address bar when you opened ComfyUI
Step 5:
- Press the 'Test ComfyUI Path and WebSocket' button on the easy installer
Step 6:
- It should now give you the option to download your Chosen Models. I am personally using the following:
- Stable Cascade 3.6B
- Resonance Prototypes Delta + Epsilon
- Refiner Lite
- Required Encoder/Decoder Models
- "Start Download!"
Step 7:
- Wait while it downloads, you fiends
It will show in the cmd.exe window for the GUI installer what it is doing. For example: 'Downloading CLIP. Finished downloading CLIP. Downloading base (Stage C) Models', etc.
In Windows you can also check your Task Manager > Performance > Wi-fi or Ethernet Network to see that it do be downloading
Step 8:
- Once it has finished downloading everything, you may close the installer
Step 9:
- Open the CAPGUI, start_GUI batch file. The cmd window should prompt you to restart ComfyUI, do so
Step 10:
- If all was successful, you should be able to access the CAPGUI by going to IP Address:6969 or http://localhost:6969/
After Install, Gen Settings
Once you have the CAPGUI installed and opened, it should be set as follows:
- Cascade Base Model: cascade/stage_c/reso_proto_epsilon-e5.safetensors
- Cascade Text Model: cascade/reso_proto_delta_e5_te.safetensors
- Cascade Refiner Model: cascade/stage_b/stage_b_lite_bf16.safetensors
- Base steps: 20 (uses Euler A as a sampler by default)
- CFG: 4
Cascade is capable of genning much higher native resolutions than SDXL, so don't be afraid to try cranking the resolution up and see what works for you.
COMPRESSION: The auto compression finder is sometimes a bit lower than what I find works best for the image. For example, at 1504x1504 resolution, it recommends 32 compression - but I find this often gives me extra limbs, wonky hands, so on. As a result, I prefer somewhere around 36-42 compression for most artists. You may want to play around with this to find the 'sweet spot' that works for your image.
After Install, Proompting
- The prototype delivers the best results when you are prompting an artist or artist mixes
- Artists should be prepended with 'by'. Ie. 'by ishikei'
- Using "by novelai" helps add contrast, amplify backgrounds and assist with anatomy when used in combination with artists
- You do not need score or quality tags, nor schizo negatives
- You want to prompt in the form of booru tags without an underscore. Ie. '1girl, standing, blue eyes, blonde hair'
- To check for artists, characters and terms the prototype dataset has learned, look at the tags .csv in the resonancemodel rentry https://rentry.org/resonancemodel/
- The higher the number beside the term, the more likely it is able to prompt it successfully - however, Cascade has shown a strong learning rate even with low hits in many cases.
-
Prompt adherence on the model is very stronk, so if you want something to show up in your image, make sure you include it in your prompt
Example prompt you can test:
anila \(granblue fantasy\), 1girl, blonde hair, large breasts, cleavage, curled horns, horns, breasts, sheep horns, short eyebrows, thick eyebrows, yellow eyes, nipples, cum, tan, tanlines, cum on body, bukkake, by blade_(galaxist), by novelai
My result for comparison - https://files.catbox.moe/87knjg.png
[ settings I used: txt2img, no negs, gen'd at 1504x1504, euler a, 20 steps, cfg 5, compression 42 ]