Resample Sigmas

Resample Sigmas takes a sigma schedule and reinterprets it for a different number of steps. You hand it a SIGMAS tensor and a target step count; it returns a new schedule with that many intervals, linearly interpolated along the curve the original points defined. The endpoints - your starting noise level and the final 0 - stay exactly where they were.

Why would you want that? Say you found a great workflow whose sampler is tuned to a 20-step schedule, but you want a quick 8-step preview while you iterate on a prompt. Normally you'd change the step count in the sampler and hope the scheduler you picked reshapes sensibly. With this node you can grab the proven 20-step schedule, resample it to 8, and feed the result into SamplerCustomAdvanced - the shape of the original curve survives at lower resolution. Same logic in reverse: stretch a schedule to more steps than it started with, for a finer pass over the same trajectory.

The mechanism is honest linear interpolation. Input points are treated as a curve, sampled at steps + 1 evenly spaced positions. That detail matters: steps is the number of intervals, and the output is one entry longer than it, because a schedule needs both endpoints. If the source schedule is a single scalar or a one-point tensor, the output just repeats that value for every timestep.

Inputs: sigmas (SIGMAS) and steps (INT, default 10, minimum 1). Output: one SIGMAS tensor ready for a sampler's sigmas input.

The thing to keep in mind is that interpolation is a lossy copy, not a recompute. A Karras schedule, for instance, concentrates steps in the middle of the trajectory; resampling it linearly smooths that concentration out, and at big step-count changes the resampled curve isn't the same sampling experience the original scheduler would give you at that step count. For a quick preview that's usually fine - and honestly the cleanest mental model is "same schedule, fewer or more samples along it." If you're chasing a specific sampler+scheduler interaction, re-run the actual scheduler at the new step count instead.

Installing it

Part of ComfyUI-ImmacTools by Immac - MIT-licensed, personal-use, no dependencies beyond ComfyUI's own numpy and torch, no model files. Install via ComfyUI Manager by repo URL or:

cd ComfyUI/custom_nodes
git clone https://github.com/Immac/ComfyUI-ImmacTools

Restart ComfyUI, then look under Example for "Resample Sigmas". Keep ComfyUI current - the pack registers through the newer ComfyExtension API.

Common issues

Don't expect the resampled schedule to reproduce what the original scheduler would emit at the new step count - interpolation preserves the curve, not the scheduler's step-placement logic. And as with the other sigma nodes in this pack, both tensor device and dtype should match before the node touches them, or torch will complain. Other than that it's a forgiving little node: pass it None and it passes None back out.