Gaussian Tail Scheduler

A different shape for the same old curve

Most of the pack's schedulers are variations on a weighted blend between sigmax and sigmin. The Gaussian Tail Scheduler from sigmas_tools_and_the_golden_scheduler is the odd one out: it builds its curve from the cumulative distribution of a normal (Gaussian) distribution, so the schedule is shaped like the tail of a bell curve rather than a power law.

Concretely, for each step the node computes

(sigmax - sigmin) * 2 * (1 - norm.cdf((x/(steps-1)) * 3.2905)) + sigmin

The magic constant 3.2905 is roughly the 99.95th percentile of the standard normal, which means the curve spans the full range: at the first step the CDF is 0.5, the factor is 1, and you start at sigmax; by the last step the CDF is ~0.9995, the factor drops to ~0.1%, and you land on sigmin. Between those ends it holds the noise level up high for a long stretch and then drops off smoothly at the very end - a slow, lazy tail.

What that means in practice: it's the opposite personality of Karras. Karras concentrates denoising effort in the middle; this one lingers near the high-noise end, so early steps do less work per step and the curve gets steeper only toward the end. That's a "gentle open, fast finish" shape that some samplers take to well.

The honest take

This is a curiosity, and it's honest to say so. Inputs are just model and steps (default 20), with a single SIGMAS output - no knobs, no model_type, no clever toggles. The README gives it one line and no comparison renders, unlike the Golden Scheduler's full gallery. If you're trying to get away from Karras's middle-heavy push on an SD 1.5 or SDXL model, it's worth one experiment on a fixed seed to feel the difference. It's also a genuinely nice teaching tool for what a scheduler does: render the same prompt with simple, Karras, and this, and the curve shapes explain the output differences almost immediately.

Like every schedule-reshaping node in this pack, it's aimed at DDPM-style models. Flow-matching checkpoints (Flux, Z-Image, etc.) want conservative, barely-reshaped schedules - a bell-curve tail is not that, so leave it alone there.

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 model downloads. This node specifically pulls in scipy.stats.norm at module load, and here's the catch: the pack's requirements.txt only lists asteval - scipy and matplotlib are undeclared. If the whole pack fails to import, that's why; pip install matplotlib scipy asteval sorts it.