Aligned Scheduler

Fewer steps, smarter placement

NVIDIA's Align Your Steps is a small but clever idea: at low step counts, where you place each step on the noise curve matters as much as the curve itself. The paper publishes hand-tuned timestep tables for common model families, and Aligned Scheduler from sigmas_tools_and_the_golden_scheduler is a ComfyUI node that turns those tables into a real SIGMAS schedule. If you're squeezing an SDXL render down to 10–15 steps, this is a legitimately useful tool.

What it does under the hood

The node takes your model, a steps count, a model_type, and an force_sigma_min toggle. Internally it looks up the 10 canonical Align Your Steps timestep indices for your chosen family, inverts them, grabs the sigmas at those points from a 1000-step "simple" schedule, then log-linearly interpolates to however many steps you asked for. That interpolation is the part that makes it flexible - the paper's tables are for ~10 steps, but you get a decent schedule at 10, 20, or 30.

The model_type dropdown (SD1, SDXL, SVD) matters, so set it to match the checkpoint you loaded. If you feed an SDXL model but leave it on SD1, the step placement will be wrong.

The force_sigma_min toggle is the interesting one. Off, with 10 steps, you get values matching ComfyUI's own built-in aligned implementation. On, the schedule adds an extra step that lands exactly on the model's true minimum sigma. The README makes a specific pitch here: unlike the paper, this node derives everything from your model's actual min/max sigmas rather than the published values, which it argues can help with COSXL models. Worth testing rather than taking on faith.

When to reach for it

This is a low-step-count play. At 20+ steps on SD 1.5 or SDXL, DPM++ 2M Karras is still the comfortable default and Align Your Steps buys you less. Where it earns its keep: SD1/SDXL/SVD workflows targeting ~10–15 steps where you want quality closer to what more steps would give you. Same architecture caveat applies as with any schedule-reshaping node - this family of tricks is built for DDPM-style models, not flow-matching ones.

Inputs: model, steps (default 10), model_type, force_sigma_min. Output: one SIGMAS, wired into a KSamplerAdvanced's sigmas input like any external scheduler.

Install & gotchas

ComfyUI Manager → search sigmas_tools_and_the_golden_scheduler, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Extraltodeus/sigmas_tools_and_the_golden_scheduler

Restart, no weights needed. If the pack won't load, requirements.txt only lists asteval while the code imports matplotlib and scipy - pip install matplotlib scipy asteval fixes a minimal install. And yes, the classic mismatch: make sure the sampler's own step count matches this node's steps, or the schedule length won't line up.