Sigmaboy Turbo 16

What it is

Sigmaboy Turbo 16 is the sixteen-step half of a two-node idea. Same pack, same one-slider UI, same trick: instead of choosing a scheduler from ComfyUI's dropdown, you hand the sampler a hand-tuned sigma curve. The node's own description points it at "16-step models (HiDream Fast, etc)" - the distilled checkpoints that land somewhere between a four-step Schnell-style model and a full 30-step render.

That in-between zone is exactly where stock schedules get awkward. A 16-step distilled model was trained to make sixteen specific-sized jumps through the noise, so the step count has to match and the spacing has to look roughly like what it saw during distillation. Generic presets redistribute effort - Karras pulls denoising into the middle, exponential front-loads it - and on a flow-matching architecture that redistribution distorts rather than corrects. A tuned curve is the shortcut around that argument.

Same caveat as its 8-step sibling: one author, October 2026, no community track record. The curves are someone's eye, not a benchmark.

How it works

Seventeen numbers: sixteen sigma values plus the terminal zero. Three stored profiles - SOFT, BASE, HARD - and a linear blend driven by the single widget.

BASE 16-step tail: ... 0.9936, 0.9885, 0.9788, 0.9600, 0.9230, 0.8550, 0.7300, 0.5000, 0.0
HARD 16-step tail: ... 0.9600, 0.9250, 0.8650, 0.7700, 0.6300, 0.4500, 0.2500, 0.0800, 0.0

Read the second-to-last value: BASE starts its final step at sigma 0.5, HARD at 0.08. So "more aggressive" means the curve dumps noise faster early and lets the last few steps work at low sigma - composition settles sooner, the tail becomes a detailing pass. On sixteen steps you have more room for that than on eight, which is the entire reason this node exists separately.

Interpolation is identical to the 8-step node: 1.0 returns BASE exactly, 3.0 returns HARD exactly, straight blend between. A SOFT profile and a below-1.0 branch exist in the source, but the 1.0 floor puts them out of reach.

Inputs, output, and the one trap that matters

aggression (1.00–3.00) is the only real control. node_id_hidden is a hidden frontend field used to route a status event back to the node's own UI - leave it alone. Output is a single sigmas (SIGMAS), wired into SamplerCustom/SamplerCustomAdvanced or in place of BasicScheduler. KSampler has no SIGMAS socket, so it can't use this at all.

Then the trap: the sigmas array's length is the step count. Wire this into a sampler and you get sixteen steps, full stop - whatever the steps widget says. Feed it to an eight-step Turbo LoRA and you've just made the model denoise in sixteen small jumps through a curve it wasn't distilled for. The result isn't an error message. It's a slower generation that looks subtly wrong - softer, or grainier, or fine in the thumbnail and mushy at 100%. That's the classic silent failure with sigma-curve nodes, and both nodes look nearly identical on the canvas, so grabbing the wrong one from the search box is easy. The 8-step node has a blue slider accent; this one is purple, and the title bar says "Sigmaboy Turbo 16". Check before you complain.

The reverse also holds: don't point it at a 50-step full checkpoint. A sixteen-value curve ending at sigma 0.5 hands a non-distilled model far fewer, larger steps than it wants.

Install

Same pack, so installing this gives you all three nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/Rimor-dev/ComfyUI-Sigmaboy

Restart, and it's under Rimor_dev/schedulers in the add-node menu. Or search Sigmaboy in ComfyUI Manager. Zero Python dependencies - no requirements.txt, dependencies = [] in the pyproject - and no model files. Bring your own checkpoint and distilled LoRA.

Where it bites

"IMPORT FAILED" in the console at startup. Built on the V3 node API (comfy_api.latest), so an older ComfyUI can't import it and the pack silently vanishes from the menu. Update ComfyUI.

No slider, just a number field. The UI comes from web/sigmaboy_16.js; hard-refresh and restart. The node computes fine with the plain widget.

Comparing it to a plain simple scheduler and seeing almost nothing. Both profiles stay close to 1.0 until roughly the midpoint, so at the 1.0 setting the difference is concentrated in the last few steps. Push to 1.3–1.5 for a comparison you can actually see, then dial back.

Console output is your friend. Every run prints the whole schedule in green, sixteen steps plus the terminal zero. If the list doesn't have seventeen entries, you're looking at the wrong node.