===========================================================================
RES-M - Think of this as a flip-side of A. Clearer and merged from more capable counterparts of models to come, including clearer anatomy and concepts. Thanks for all the support, more to come. Please share your generations!
RES-X - A few editions coming soon now that FLUX is a primary focus. Plan to release the "final contenders" in the coming week or so (9/14)
RES-A - The 'A' is for analog. It's a richer, softer, more photo-realistic take on Stable Cascade
R35 models are created by DR34MSC4PE with all the trappings you've come to expect as well as some cool bonuses:
Generate stunning, enormous images quickly or "Analog" poloroids in a snap. The model's compression dial allows for whichever poison you prefer!
NSFW tuned and capable of realism and artistic/anime
Full BF16 C/B stages with functional workflows allow for quick, easy setup in Comfy UI. No managing unets/CLIP!
Users who prefer Diffusers version of the models can claim them in the coming weeks!
Comfy nodes (like those found in some of the example images) that affect CLIP for SDXL work wonderously on Cascade as well. LORA support demonstrated and disabled therein.
Support for proper use in Automatic/SDNEXT - Stay Tuned!
Limitations:
Mid range detail, eyes can be inconsistent/different-colors without adjusting resolution/compression/steps accordingly
Can gravitate toward animated output at low step/compression
Nude male anatomy (WIP)
Blurryness in some NSFW concepts at low step counts (WIP)
===========================================================================
DR34MSC4PE is
c0ur4ge (trainer/qa/inference code) /
eraser851 (captions/tooling code/data/qa) /
===========================================================================
Usage Guide - For all R35ON4NCE Variants:
Use with the default Cascade Workflow or the advanced found in the demo images.
Examples (pastebin):
===========================================================================
C-Stage
===========================================================================
CFG: 3-8
Latent Compression: 22-42
Steps: >20-80
Note: higher steps can reduce blur in mid-range details in conjunction with compression settings. The interplay between these settings is the difference between mush and miracles.
CFG works differently in Cascade. While it can improve prompt coherence it can impact (or improve) image quality in unexpected ways. Lower CFG can often yield more detailed images, for example.
===========================================================================
Description
This is the A-stage (VQGAN VAE) component of Stable Cascade. It is packaged in with the Comfy C-stage checkpoints automatically and does not need to be download except for UNET style inference or within the advanced workflows. It produces crisper results then the A-stage checkpoint released with Cascade after fine tuning on HQ images by madebyollin.
FAQ
Comments (9)
Thanks. Where can we find a workflow for this?
Most of the images have them embedded and you can simply click the "More information" for lots more info about recommended Samplers/pastebin links to the json workflows therein as well!
Expirimenting with Samplers/Compression workflows is half the fun/modular nature of Cascade! Have fun and holler if you have questions.
@c0ur4ge
Thanks for your reply!!
@c0ur4ge
Probably a stupid question, but where is the more info button for images ?
@strong134 So, save the image and load it into comfy, as the workflow is embedded within the image itself. Some of the prompts are from the OneButtonPrompt node, so loading them and clicking generate (after (updating the checkpoints to the downloaded versions in Comfy) should let you see the prompt in the console. Else, click "More Info" above and scroll down to download the .json files from pastebin to load directly into comfy, sans image!
That's to say - the information's coming from INSIDE the image. ;)
@c0ur4ge
Go it. Thanks!
can it use with ultrapix,tks
Yes, eventuall! These are Comfy Checkpoints (which glue together some of stages) - I would argue that the advance upscale workflow found here (with 2x nearest latent upscale) accomplishes similar results to Ultrapixel but support will be added along with the next R35 C-stage release + HuggingFace Diffusers checkpoints.
Just an update here - this has been tested and is working. Looking to finish up a new NSFW checkpoint and the UC diffusers release of them in the coming week. Thanks for your patience!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.