An update to my previous post on Stable Cascade with ComfyUI - instead of requiring four separate model files, we now only need two checkpoints, and the ComfyUI workflow is now very straightfoward!


Download the stage b and c checkpoints to ComfyUI/models/checkpoints.

And update ComfyUI. That’s it.


Here is something I was playing around with, since ControlNets / LoRAs have not been trained to work with Stable Cascade yet. I wanted to use Stable Cascade to covert monochrome line art (canny) into something photo-realistic:

Stable Cascade merged checkpoints for ComfyUI

  • Load the Stage C checkpoints,
  • Create prompts as usual using the embedded CLIP from the Stage C checkpoint,
  • Create a compressed latent space with StableCascade_EmptyLatentImage,
  • And wire-up all this the first Stage C KSampler node.
  • Then, wire-up the positive prompt and the latent output to StableCascade_StageB_Conditioning (no zero-ing out required).
  • You know the drill for the Stage B KSampler node...
  • Now, use the embedded VAE (Stage A) from the Stage B checkpoint to get the line art image.
  • For the next step, first, encode the image with StableCascade_StageC_VAEEncode,
  • And use the output latents in a second pass through the Stable Cascade model.
  • I adjust the prompts and most importantly, the denoise value in the Stage C KSampler such that its as low as possible (to look as close as possible to the input line art image) while producing a more photo-like image.
    • Adjust just a little lower and you will get a colored line art drawing, if that is what you prefer.
    • A little higher will be photo-realistic but the composition will differ slightly.