LogNormal Scheduler

The LogNormal Scheduler is the sibling of GaussianScheduler, and it exists because noise doesn't behave like a plain symmetric bell curve. A log-normal distribution is a bell that's been skewed - a long tail stretching off to one side, the mass of it concentrated at the other. Applied to a sigma schedule, that gives you a curve where most of the interesting denoising happens in a concentrated region and the rest of the run trails off gently. It's the sort of shape people hand-build when they want a middle-heavy schedule with soft landings on both ends.

How it works

Like its Gaussian sibling, it evaluates a probability density function across an x-window defined by a and b, then rescales the curve to sit exactly between sigma_min and sigma_max. The difference is the distribution: instead of a symmetric bump, you get the log-normal's distinctive asymmetry. mean shifts where the peak of that skewed bump lands, and standard_deviation controls how spread out (and how lopsided) it is. Small stdev gives a sharp, heavily-peaked curve with most steps doing almost nothing; larger stdev spreads the effort into a gentler dome. Either way the shape is asymmetric in a way a plain Gaussian isn't, and that asymmetry is the point.

The inputs that matter

• steps (20) - number of sigma values.
• sigma_max (1.0) / sigma_min (0.0) - endpoints. 0–1 by default.
• mean (0.0) - where the skewed peak sits.
• standard_deviation (1.0) - spread and lopsidedness. Your main shape dial.
• a (0.0) / b (1.0) - the x-window. Swapped automatically if a > b; an error if they're equal.

Output is a SIGMAS array.

Install

Part of ComfyUI-ScheduledGuider-Ext. ComfyUI Manager - search ComfyUI-ScheduledGuider-Ext - or:

cd ComfyUI/custom_nodes
git clone https://github.com/mfg637/ComfyUI-ScheduledGuider-Ext

Restart ComfyUI. No models, no extra dependencies.

Common issues

By now you know the drill: default range is 0–1, which is ideal for a guider's CFG curve and wrong for a raw sampler schedule - run it through ScaleToRange first if it's going to a KSampler.

The subtler thing: a log-normal's long tail means some of your sigma values sit very close together, which is fine for a CFG curve (denser points = smoother interpolation) but can waste sampler steps if you feed it as a real schedule on a model that doesn't want concentration. And the flow-matching caveat from GaussianScheduler applies double here - heavily skewed schedules are exactly the kind of aggressive reshaping that flow-matching models dislike. SD 1.5/SDXL, or a guider curve, is where this node belongs.