ModelSamplingStableCascade

Be honest about why you're here: you downloaded a Stable Cascade checkpoint, or you're looking at an old workflow, and something told you this node needs to be in it. Fine. ModelSamplingStableCascade is the sampling patch for Stable Cascade's stage C - the efficient two-stage model Stability released in early 2024 and then promptly stopped talking about as SDXL fine-tunes and later SD3 kept rolling out. It's not the worst model ever shipped; it just never won, and by 2026 almost nobody generates with it.

The node itself is about as simple as the family gets. One required input - shift, default 2.0 - and a MODEL output. It clones your model and swaps in a StableCascadeSampling object, which is a discrete cosine-schedule sampler with the shift applied. In source terms, shift skews the log-SNR curve of the noise schedule: logSNR += 2 * log(1/shift), which rebalances how much denoising effort lands on the early composition-heavy steps versus the late detail steps. Crank it up and the model spends more time shaping structure; drop it toward 1 and you approach the raw cosine schedule.

When you'd actually reach for it

If you're running Stable Cascade at all, you want this node in the workflow right after the loader, and the default of 2.0 is a reasonable starting point - it's the value the official workflow shipped with. The practical advice from the model's short community life: generate the base at a lower resolution and let stage B upscale, and keep the shift modest. It's not a dial that rewards being pushed hard, and the model was never popular enough that there's a rich lore of what happens at shift 8. Nobody is tuning this.

The honest read

This node landed in February 2024, two weeks after Stable Cascade's release, and it's a perfect snapshot of its era: every flow or cosine-schedule model got a bespoke sampling node because the model's own config didn't always carry the right numbers. Now the same shape of node exists for Flux, LTX, and AuraFlow - the difference is those models are still being used. If your Stable Cascade workflow isn't producing, the sampling patch is probably not where the problem is; the problem is usually that the model itself wants more care (and more VRAM via the C + B two-stage setup) than it gives back. Use it, but don't expect it to resurrect the model.