Selector In (SIGMAS)

SelInSigmas ("Selector In (SIGMAS)") routes SIGMAS objects by model type: eight sigma inputs, one model_type integer, one SIGMAS output. If you're not deep in ComfyUI's advanced sampling, that sentence already means nothing to you - so let's back up. SIGMAS is the data type that carries a noise schedule: the list of noise levels the sampler steps through. It's what sits between a BasicScheduler/KarrasScheduler and a SamplerCustom in advanced-sampling graphs, and it's the knob that determines whether your model gets a Karras-shaped, exponential, or custom-shaped denoising trajectory.

Why route it per model? Because the "right" schedule genuinely depends on the model family. The classic SDXL default is a Karras schedule; flow-matching models like Flux and AuraFlow dislike aggressive schedule reshaping and want conservative, balanced sigma curves - Karras can actively hurt there. Guidance-distilled models are pickiest of all. If one workflow carries both an SDXL branch and a flow-matching branch, the sigma schedule that's perfect for one is wrong for the other. SelInSigmas lets you pre-wire a KarrasScheduler on port 1 and a BasicScheduler on port 3 and let the model_type broadcast choose between them.

How it works

Pack standard: model_type (1–8) selects sigmas<number>, winner out the single SIGMAS output into a SamplerCustom. Ports are lazy, so only the selected schedule branch evaluates - worth noting, since a KarrasScheduler call is cheap but the downstream sampler run it controls is anything but. Everything stays in step because it's the same model_type number that's selecting your model and your steps elsewhere.

Inputs & outputs that matter

• model_type (INT, required) - the selector, normally fed by RecourseCkpt.
• sigmas1 … sigmas8 (SIGMAS, optional) - candidate schedules from BasicScheduler, KarrasScheduler, ExponentialScheduler, and friends (ComfyUI-Advanced or the built-in advanced sampling nodes).
• SIGMAS output - the winner, into SamplerCustom.

Installing it

The pack standard. ComfyUI Manager → search comfyui-selector → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/exdysa/comfyui-selector

Restart and it's under Selector_Recourse/In. No requirements.txt, no model files, no network access. You'll also need the advanced sampling nodes that actually produce SIGMAS objects.

Gotchas

• Nothing here produces a schedule. The SIGMAS objects come from scheduler nodes; this node only chooses between them. If you never see SIGMAS outputs in your graph, you're not in advanced-sampling territory yet - and that's fine, the plain KSampler doesn't use this.
• Empty selected port → None out → the sampler won't run. Check your model_type number first.
• Sigma schedules are the subtle end of the tuning spectrum: picking wrong isn't a crash, it's subtly worse images or slower convergence. Don't wire this in just to have it - only when you genuinely maintain per-family schedules.
• Small single-author pack (GPL-3.0), quiet since early 2025. The niche is real but narrow; for one model family, a single scheduler is all you need.