Type | Workflows |
Stats | 516 |
Reviews | (58) |
Published | Apr 20, 2024 |
Base Model | |
Hash | AutoV2 92A99F9683 |
Important Note:
This LCM Workflow is highly experimental, not all prompts will work correctly, generations will break - this is normal, workflow is designed to fight such problems at some extend. Treat this workflow more like a new way of generating pictures that is different from SDXL and SD1.5, explore new ways of using Würstchen v3 architecture.
How to Install:
Download the Official Stable Cascade *.safetensor files at:
Move Stage B and Stage C to - ComfyUI\models\checkpoints
Move Canny to - ComfyUI\models\controlnetUnzip downloaded "The Ravens Workflow v01.zip" archive.
I'm using sd-perturbed-attention custom node, you can use mine and move folder "sd-perturbed-attention-master" to - ComfyUI\custom_nodes, also you can download it from the official github page or install any other way you see fit.Move "The Ravens Workflow.png" inside ComfyUI window.
Documentation:
Würstchen v3 aka Stable Cascade
"The Ravens" Workflow
v 0.1 Impressive and Soulless
It is a bit hard to wrap all the knowledge I've gained over the past months about the Würstchen v3 architecture and my experience with it in this short text window. So, right now I'll concentrate on my general ideas, practical tips and trick. I made my best to make this workflow clean and easy to use.
1. CONCEPTS:
My whole idea is revolving around LCM Sampler and how Cascade's latent noise "deviate" with it. LCM Sampler creates a geometrical noise patterns, this patterns help generations to achieve clean geometrical shapes, symmetrical harmony in picture, better details even at low resolutions, textures have correct shapes, lines are clean and straight almost like vectors.
2. PROBLEMS:
But there is a problems - some generations became broken with certain prompts, there is different artifacts: dots, lines, white spots, distortion and noise patterns - to battle this I made this 3 pass workflow: "Main Pass, HiRes Fix First Pass, HiRes Fix Second Pass". Usually after whole process all artifacts became more "hidden" or vanish completely. Check all 3 generated pictures to find the best one, since the last picture may be too smooth.
3. SAMPLERS:
Two additional HiRes passes not really for getting high resolutions, but to battle problems. Combining different samplers and schedulers at all 3 stages help different prompts to work properly.
LCM-Simple - This is your main, always try to use it at all stages, especially Hi-Res First and Second Pass.
Euler-Normal - If there will be heavy artifacts in the picture, then it is time to use this sampler. first use it at "Stage C" of "Main Pass" - check if it helps, if not then at "Stage B" of "Main Pass". Then continue to change samplers in chain like fashion at HiRes Passes.
Euler-A-Karras - Our artifacts free friend, but almost without the effect of LCM Sampler. Use it to fix really noisy pictures, start with "Stage C" of "Main Pass" and then move along in the same chain-like fashion to find the sweet spot.
Euler-A-Simple - if nothing helps, and picture is a mess and not react to prompt properly, then it is time to get back to this old vanilla cascade sampler, usually using it at "Stage C" of "Main Pass" will fix the picture, if not - you know what to do.
You can try different schedulers for samplers, they all help with different stuff.
4. RESOLUTION \ COMPRESSION:
Resolution and compression factor is a funky beast. It not just fix the "double heads" problems, but also can change overall composition and details - for example pixelart with compression of 32 and resolution 1024x1024 will create smaller "pixels" then 64 compression and will change the style.
Compression 32 - Works from 768x768 to 1280x1280 at all aspect ratios.
Compression 48 - Works from 1280x1280 to 1792x1792 at all aspect ratios.
Compression 64 - Works from 1792x1792 to 2304x2304 at all aspect ratios.
Using 64 compression at lower resolution and different aspect ratios can give an interesting results. Usually there is no real point to change compression at "HiRes Fix Passes" since we always want that 32 crisp look.
5. STEPS:
The more steps the better, as always, but there is a sweet-spot:
Main Pass - Stage C 12-36+ (Sweet-spot is 18).
All HiRes Passes - Stage C 6-18+ (Sweet-spot is 9).
All Stages B - 6-12+ (Sweet-spot is 6).
Don't forget that if you rise CFG scale too high you also need to rise steps.
6. CFG:
The higher the CFG scale, the less variation there will be in the picture, they starts to burn and become more bright, if you want dark picture - lower the CFG scale. Your prompt will get a noticeable boost at high CFG numbers.
Stage C - 0.8-18+ (Sweet-spot is 2.5).
Stage B - Leave it at 1.5.
7. DENOISE STRENGTH:
For some reason cascade love to add a lot of noise, so sticking to 0.1 strength value is a good idea. If you need smaller changes you may use 0.05, but if you want heavy changes you may use 0.2 or 0.3+. For our fixing purposes sweet-spot is 0.1 value.
8. WEIGHTS:
Right now weights for tokens don't work, avoid (this kind:1.0) of formatting - you will get only burned images in most cases, just move tokens at the start of prompt to increase weight.
9. PAG:
This handy tool is used to affect generations and make them more focused on details. But I like to use it to add variations to the same seed, by tweaking "adaptive scale", and "scale" parameters.
10. CONTACTS:
You can contact me at civitai.com/user/Temp or reddit.com/user/-Ellary-
Now scroll up and find the workflow, run the demo prompt. Godspeed o/
FAQ:
-How can I help?
:Spread the word about Stable Cascade and Würstchen architecture, show us your best gens at CivitAI and Reddit, teach others how to use Stable Cascade.
-One of the "HiRes Fix Passes" generates blurry pictures!
:Play with "scale by" numbers, noise must be aligned.
-I like the generated picture but it is too dark!
:Rise the CFG for one\both "HiRes Fix Passes".
-How can I disable different modules?
:"HiRes Fix VAE ENC - R1" and "R2" - to disable HiRes Passes.
"Image Size" - to disable Main Pass.
"APPLY CN" - to disable Control-Net Canny.
"CLIP-V CON - I1" and "I2" - to disable Clip-Vision.
-How to use Img2Img?
:1. Load your image to "Load Image I1".
2. "Upscale Img2Img - to R1" connect to "HiRes Fix VAE ENC - R1".
3. Usually you want to disable "Main Pass" when you use Img2Img.
-How to use Control-Net Canny?
:1. Load your image to "Load Image I1".
2. Enable "APPLY CN".
-How to use Clip-Vision?
:1. Load your image to "Load Image I1" or "Load Image I2".
2. Enable "CLIP-V CON - I1" or "CLIP-V CON - I2".
-How much VRAM do I need?
:8GB VRAM, 16GB RAM is enough.