Schedule Resample

The Schedule Resample node takes a verified Sigmax schedule and rebuilds it at a different step count via explicit normalized-index interpolation. Plain English: you have a 50-step schedule and you want the same curve sampled at 20 steps, or vice versa - same shape, different resolution. It's the "change how many steps this recipe runs at" node, and it exists because that operation is easy to do badly and the pack insists on doing it traceably.

This is genuinely useful once you understand the trap it's avoiding. A schedule's shape is a statement about how denoising effort is distributed; resampling isn't "pick every Nth number," because that throws away the shape. Normalized-index interpolation re-expresses the curve on a new step grid while preserving the curve. That's the difference between a schedule that still behaves like the recipe and one that silently becomes a different recipe.

Inputs and outputs

• sigmas - the SIGMAS from a Sigmax scheduler (or from an earlier Slice/Concatenate/Resample).
• schedule_info - the accompanying JSON.
• output_steps - where you want the new count, default 20.

Outputs: sigmas and a fresh schedule_info. Crucially, the metadata travels with the edit - the result is still a verified Sigmax schedule the inspection family can check, which is the point of doing this inside the pack rather than with a raw tensor splice.

Installing

Standard Sigmax install:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-Sigmax comfyui-sigmax

Restart, search Sigmax. ComfyUI Manager has "ComfyUI-Sigmax". Python 3.10+, ComfyUI 0.29.0+, no extra dependencies.

Where people get burned

The main trap is expectation management: resampling a schedule does not make a model "want" fewer steps. If you resample a 50-step RAW recipe to 8 steps, you get a curve that looks like the RAW curve on a short grid - but RAW is a 50-step model, and no interpolation fixes that. Use this for matching step counts within a recipe's plausible range, not for forcing a step count a model wasn't designed for. And the standing rule applies: the resampled sigmas feeds the custom-sampling path directly, no second scheduler, no extra shift. When in doubt, run the ScheduleInspector on the result to confirm it stayed coherent.