What are the stages in Stable Cascade?

Stable Cascade has three stages. Stage C is the text-conditioned prior that does most of the creative work in a highly compressed latent space. Stage B decodes that into a larger latent, and Stage A is a lightweight VAE that produces the final pixels. You mainly tune Stage C and Stage B.

How many steps should each stage use?

Stage C typically needs 20 steps for balanced quality (10 for drafts, 30+ for maximum), while Stage B needs far fewer — around 10 steps is plenty because it is only decoding, not generating. More Stage B steps give diminishing returns.

What guidance scale works best?

Stage C responds well to a CFG of around 4, lower than typical SD models because its compressed latent space is very prompt-sensitive. Stage B usually runs at CFG 1.1 or even 0, since it is decoding rather than interpreting the prompt.

Why is Stable Cascade so memory efficient?

It works in a 42x compressed latent (Würstchen architecture), so the expensive Stage C operates on tiny tensors. That lets it generate high-resolution images with less VRAM and fewer steps than pixel- or 8x-latent diffusion models.

Stable Cascade Configuration Guide

Stable Cascade configuration

Stable Cascade (the Würstchen v3 architecture) splits generation across three stages working in a heavily compressed latent space. Stage C is the text-conditioned prior doing the creative work, Stage B decodes it into a larger latent, and Stage A is a tiny VAE producing pixels. Most of your tuning happens on Stage C and Stage B, and this guide picks sensible values for your quality target and VRAM.

How it works

Because Stage C operates on a 42x-compressed latent, it needs surprisingly few steps and a low guidance scale to produce strong results. Stage B mostly decodes, so it needs even fewer steps and almost no guidance. The tool maps a draft/balanced/maximum target to step counts for each stage, suggests a CFG pair, and recommends a latent resolution that fits your GPU — warning you when the full bf16 pipeline is tight on smaller cards.

Tips and notes

Don’t over-step Stage B. Beyond ~10 steps it adds time without real quality gains; spend your budget on Stage C instead.
Keep Stage C guidance low. A CFG of ~4 is the sweet spot; high CFG over-saturates and distorts in the compressed latent.
Use bf16. Stable Cascade ships in bfloat16 weights; running fp32 doubles VRAM for no quality benefit.
Square resolutions are safest at 1024x1024; push to 1536 only with ample VRAM, as the prior was trained around 1024.