Hangover Sigmas Function

The scheduler dropdown is a black box with 16 names you didn't choose - normal, karras, sgm_uniform, beta, linear_quadratic, and whatever a workflow from Reddit demanded. Hangover Sigmas Function replaces the dropdown with a formula. You pick a curve (or write your own expression in x and p), pick a step count, and get a SIGMAS list you feed straight into a custom sampler.

Is that a good idea? Depends entirely on your model. On SD 1.5 and SDXL checkpoints - including Illustrious/Pony-style anime finetunes, still DDPM-style no matter what year they were merged - reshaping the noise curve is a genuine correction, which is the whole reason "Karras" got glued onto so many sampler names. On flow-matching models the trajectory is already near-straight, so a reshaped curve doesn't correct anything, it distorts: the community's 62-sampler × 16-scheduler sweeps keep finding Karras and exponential-style curves failing across the board there. Same node, opposite verdicts.

How it works

Under the hood it's almost embarrassingly simple, which is good: it builds a linspace, runs a function over it, and rounds every value to five decimals.

• Exponential (x^p) - linspace(1.0, 0.0, steps+1) ** p. The label oversells it; x**p is a power curve, not an exponential. p = 1 is straight linear. p > 1 front-loads the work: you plunge through the high-noise steps fast, then crawl through a long low-sigma tail. p < 1 lingers up high and drops hard at the end.
• Cosine (from 0 to pi/2) - a cos arc from 1 to 0. Rounder than linear, gentler in the middle. Note that p is ignored here; the widget is just along for the ride.
• Gauss (e^-x^2 from 0 to p) - exp(-x²) sampled from 0 to p, so p is the range, not a strength. More below, because this is the trap.
• Custom function (use x and p) - your expression, evaluated at every x between x_from and x_to, with p as a variable. Parsing is sympy, so real expressions work; a typo throws instead of silently doing something odd.

Inputs: function, steps (INT, default 8, minimum 2), p (FLOAT), custom_function (STRING, default x**p), x_from and x_to (FLOAT). Outputs: sigmas (SIGMAS) - wire it into SamplerCustom/SamplerCustomAdvanced, or anything else that eats a SIGMAS, like FlipSigmas or SplitSigmas - and steps (INT), the same number as an integer, so a downstream node that wants a step count can take it from here.

The catch: the sampler walks the list, so the list is the step count. You're getting steps + 1 values, and that's what actually sets how many denoising steps run.

The three traps

1. The default Custom expression points the wrong way. x**p with x_from = 0 and x_to = 1 gives you 0 → 1: from no noise to full noise. Sigmas must run high-to-low - either flip the range (x_from = 1, x_to = 0) or write the curve backwards with (1-x)**p.
2. Gauss with p = 1 never reaches zero. The last sigma is exp(-p²), so at the default p = 1 your schedule ends at about 0.368, and residual noise stays in the latent. Push p to about 3.5 or higher and the last value rounds to exactly 0.0; below 3.4 it doesn't.
3. The built-in curves are baked on a 1 → 0 scale. That matches flow-matching models, but not SD 1.5/SDXL, whose schedules start around 14.6 - hand an SDXL checkpoint a 0–1 schedule and you've described a late refinement, not a generation. Use Custom mode: p*(1-x) with x_from = 0, x_to = 1, p ≈ 14.6 is a plain linear SDXL-scale ramp.

Install

Manager: search ComfyUI_Hangover-Utils, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Hangover3832/ComfyUI_Hangover-Utils
pip install -r ComfyUI_Hangover-Utils/requirements.txt

That pip line isn't in the README and it isn't optional. This module imports sympy (for custom_function parsing) and matplotlib at the top, and __init__.py imports it unguarded - a missing sympy takes out the whole pack, not just this node. The pack also uses ComfyUI's newer comfy_api.latest node API, so a stale install fails at import. Update first, then debug.

Should you reach for it

For a flow-matching model, treat it as a teaching tool: hand-reshaped curves are mostly where you lose quality there, and the useful knob has moved to the model's timestep shift. For SDXL-family models at low step counts it's worth an afternoon - build x**0.5, plot it, run it, compare against Karras. Pair it with Plot Sigmas from the same pack: plot before you sample, or you're guessing with extra steps.