SamplingPercentToSigma

SamplingPercentToSigma answers one boring-sounding question that turns out to matter a lot in custom sampling: what noise level corresponds to X percent of the way through denoising? It takes your model, a percentage from 0 to 1, and returns the sigma value sitting at that point on the model's own noise schedule. It's a tiny node with no sampler in sight - the output is a plain FLOAT - but it's the glue that lets you build and reason about custom sigma pipelines instead of guessing at numbers.

How it works

The model wrapper already knows its noise curve: for any timestep it can tell you the sigma, and percent_to_sigma inverts that. Sampling percent 0 means "the very start" - pure noise - and 1 means "the end" - clean. The node interpolates the model's sigma curve at the requested fraction and returns that sigma as a float.

Inputs:

• model - any MODEL. Important: the answer is schedule-specific. 70% on an SD 1.5 schedule is not 70% on a Flux schedule, because each model has its own sigma curve.
• sampling_percent (default 0, 0-1) - where to look.
• return_actual_sigma (default false) - and here's the subtlety. The node's tooltip spells it out: at exactly 0% it returns a sentinel 999999999.9, and at exactly 100% it returns 0.0 - the values used for interval checks, not the real sigmas. Flip this boolean on and you get the actual sigma_max / sigma_min at those two endpoints instead.

Output: sigma_value, a FLOAT.

Why you'd reach for it

Two honest use cases. First, feeding values into ManualSigmas - you want a custom schedule that stops at "roughly 80% through", and this node gives you the exact sigma to type in. Second, driving logic in your graph: you can compare a sigma against this value and branch a workflow, which is exactly what the "interval checks" phrasing in the tooltip is about. If you've ever stared at a ManualSigmas widget wondering what a reasonable sigma even is, this is the lookup that answers you.

Where people get burned

• The sentinel isn't a bug. 999999999.9 at 0% is intentional - it's a value that compares greater than any real sigma, which is what an interval check wants. If you expected to see the model's actual max noise, flip return_actual_sigma.
• Model matters. The result is meaningless without the right model wired in. Ask SD 1.5 and you get one curve; ask Flux and you get another. Same percent, different sigma.
• It's a float, not a SIGMAS. It won't plug into SamplerCustomAdvanced's sigma input by itself - you use it to build or check schedules, usually via ManualSigmas.

Ships with ComfyUI core. Not glamorous, but the kind of node that makes advanced sampling feel less like numerology.