ModelSamplingStableCascade (Ray)

Stable Cascade was Stability's ambitious two-stage "Wurstchen" model - a compressed latent stage plus a refinement stage - that never quite became the community default it was meant to be. It still shows up in niche workflows, and if you're running one through Raylight, this is the sampling patch it needs. RayModelSamplingStableCascade is a single-knob node: shift, default 2.0.

Mechanically it's the sibling of the SD3 patch in this pack - it clones the model on every Ray worker and swaps in a StableCascadeSampling base with an EPS sampling type, then hands the ray_actors stream back unchanged. The one difference is that Stable Cascade's noise schedule has its own character, and shift compensates for it. Cascade's stages have different resolution behavior than SD3, which is why its reference shift (2) is lower than SD3's (3). Treat the model card for your specific stage as the authority; the default is the safe starting point, not a guaranteed optimum.

What it is NOT

A few things to clear up, because Cascade confuses people even on single-GPU ComfyUI:

• This node doesn't do the two-stage orchestration. You still run the base and refiner stages separately; this just fixes the sampling math for whichever model you've loaded on the workers.
• It's not an upscaler. The "cascade" name makes people assume it's a resolution ladder. It isn't - it's a noise-schedule patch.

If you're on a single GPU, ComfyUI's built-in ModelSamplingStableCascade gives you the same patch without needing Ray at all. Use the Ray version only when the model is actually loaded on Ray workers.

Install

Same as every node in the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/komikndr/raylight
cd raylight
pip install -r requirements.txt

Restart, or install "raylight" via ComfyUI Manager.

Gotchas

The honest version: this node exists mostly for completeness. Stable Cascade is a legacy model, and Raylight's own supported-models table doesn't even list it - the pack's active attention is on Wan, Flux, Qwen, Chroma, and the LTX family. If you're building a Raylight pipeline around Cascade, you're on your own for tuning, because you won't find many community workflows to crib from. It's here if you need it; just don't expect the ecosystem to hold your hand.