⭐ Star Custom Sigmas

What it's actually for

Every ComfyUI sampler builds its noise schedule from two widgets: a step count and a scheduler name. ⭐ Star Custom Sigmas skips both. You hand it the actual sigma values and the sampler uses them, unchanged, start to finish.

That sounds like a power-user toy until you remember what changed in 2026. On flow-matching models - Flux, Qwen-Image, Z-Image - the trajectory from noise to image is nearly straight, which means there's very little for a clever schedule to fix and a lot for it to break. Karras and exponential went from safe defaults to near-universal failures on those architectures. Meantime half the turbo and distilled checkpoints ship with a published, hand-tuned sigma list in the model card, because that list is what the distillation was trained against.

So there are two real reasons to reach for this node: the model card gave you a schedule, or you're tired of a scheduler eating a step you wanted and you'd rather write the numbers yourself.

How it works

It's a converter, and a small one. The sigmas text field is parsed with a regex for numbers, turned into a float tensor, and handed out as a genuine SIGMAS value. Because that's a real ComfyUI type, the output drops into any SIGMAS socket - SamplerCustomAdvanced, SplitSigmas, FlipSigmas - not just StarNodes' own sampler.

The counting rule is the thing people trip on: N values mean N−1 sampling steps. The list is the step boundaries, not the steps. The default 8-step list has nine numbers in it:

1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0.0

When you wire this into ⭐ StarSampler (Unified)'s options input, the sampler detects a tensor there, takes the step count from numel() - 1, and overrides its own steps and scheduler widgets. Its info output then reports the scheduler as custom sigmas, so you can confirm it took. The sampler_name widget still applies - euler stays euler.

The pack also bakes in four Qwen-Image 2.1 Turbo schedules (8, 10, 12, 16 steps), because those are the ones shipped as fixed lists rather than as a scheduler. The author's docs say picking a preset writes its values into the text field so you can edit from there, and that editing the field away from the preset flips the dropdown to Custom.

The inputs and output that matter

Only two things to set:

• preset - QI 2.1 Turbo- 8/10/12/16 Steps, or Custom. Any preset other than Custom ignores the text field entirely, so don't sit there editing a preset's numbers and wonder why nothing changed.
• sigmas - comma-separated, high to low. Multiline is fine, extra whitespace is fine. Only read when the preset is Custom.

Output: sigmas (SIGMAS), into a sampler's options input or any vanilla SIGMAS input.

Install

ComfyUI Manager → search Starnodes → Install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

The pack's requirements.txt pulls psd-tools, opencv-python, scikit-learn, webcolors, color-matcher, imageio-ffmpeg, requests, beautifulsoup4 and ollama - but none of that is for this node. If a package fails to build, the pack still loads and only the affected nodes disappear, which the README says outright. Star Custom Sigmas needs nothing beyond torch. Find it by double-clicking the canvas and searching star.

Where people get burned

The numbers are used verbatim. No rescaling to the model's range, no safety net. That 1.0 → 0.0 list is right for flow models; a classic SD 1.5 or SDXL checkpoint expects something starting up around 14.6 and ending at 0. Feed an SDXL model the default list and you'll get mush, then blame the node.

No two identical values next to each other. This is the classic custom-sigma failure, and it's usually 0.00, 0.00 at the tail, producing a blank image. Any duplicated adjacent pair can blow up the step.

Big jumps between neighbours cost you. A drop from 0.95 to 0.80 in one step is where the ugly grid patterns on Flux backgrounds come from. If you're hand-writing a schedule, keep the curve smooth in the high end and do your shaping in the middle.

It gets ignored, quietly-ish. Connect a detail schedule to StarSampler and the custom sigmas are dropped with a console warning. Fewer than two parseable values and it falls back to 1.0, 0.0 - a one-step run that looks like a bug.

One genuinely useful trick: you don't have to invent a schedule from scratch. Run a sampler with a scheduler you like, read the sigma values out of the console/log, paste them in, then edit the shape - push a couple of mid-list values up if faces and hands are coming out soft, or trim the tail if you want the composition without the over-baking. That's the whole workflow, and it's easier than it sounds.