AlignYourStepsScheduler

AlignYourStepsScheduler (AYS) drops in a set of pre-tuned noise levels for SD 1.5, SDXL, and SVD, and interpolates them to whatever step count you ask for. It came out of NVIDIA's "Align Your Steps" research, which found that the shape of the noise schedule matters more than the step count for few-step generation - a set of well-placed sigmas beats a generic curve when you're trying to squeeze quality out of 10 steps.

It ships with ComfyUI core (comfy_extras/nodes_align_your_steps.py). Nothing to install.

How it works

Rather than computing sigmas from a formula, it starts from a fixed table of empirically-tuned noise levels per model type - SD1, SDXL, and SVD each have their own list. The table has 11 entries, and if your steps doesn't match, the node log-linearly interpolates to hit exactly steps + 1 sigma values, then crops for denoise just like BasicScheduler does. The result is a schedule that spends denoising effort where those models actually benefit from it.

The inputs are short:

• model_type - SD1, SDXL, or SVD. Pick the family your checkpoint belongs to.
• steps - default 10. This is a few-step schedule; its whole design goal is good output at low step counts.
• denoise - 0–1, same img2img cropping behavior as elsewhere.

Output is a SIGMAS tensor for SamplerCustomAdvanced.

Where it fits in 2026

Let's be honest about its position: AYS was a big deal in the SD 1.5/SDXL few-step era, and it's mostly been superseded for newer models - Flux, Z-Image and friends either ship their own schedulers (like Ideogram 4 Scheduler) or want conservative generic schedules. Corpus mentions have been drifting down since 2025.

But it's not dead. Two places it still earns its keep:

• SDXL at few steps - an AnimateLCM or turbo-LoRA-style SDXL workflow at 4–10 steps pairs naturally with model_type=SDXL. That's the use case it was built for.
• SVD - it's one of the only schedulers with an SVD-specific preset, so video-from-image workflows built on Stable Video Diffusion still reach for it.

Where people get burned

• Using it on a model family it has no preset for. There's no Flux row, no flow-matching row. Forcing SDXL onto a 2026 model because it's "close enough" is how you get mush - use the model's own scheduler instead.
• Defaulting to 10 steps and stopping. The scheduler is tuned for few steps, but the sampler still has to cooperate; a converging sampler at 10 steps is a starting point, not a finish line.
• Expecting a speedup by itself. AYS changes where the work happens, not how much. It makes 10 steps look better; it doesn't make 20 steps faster.

The niche is small but real. If your model isn't SD1/SDXL/SVD, there's almost always a better, model-specific scheduler waiting.