ARC Scheduler

The ARC Scheduler (class ARC Scheduler) is a niche in the Switchblade Pack's generation corner: it produces a SIGMAS schedule for advanced samplers, and its one claim is being deterministic. Same model, same steps, same denoise - same sigmas, every time. That's genuinely useful when you're A/B-testing samplers and want the schedule out of the equation, and it's about all it promises. Nothing here calls an API, needs a model download, or reads your prompt; it's a math node.

How it works

The source tells the story plainly. First it builds a baseline schedule: log-space interpolation between the model's sigma_max and sigma_min (defaults 14.0 and 0.01 if the model doesn't expose them), descending high→low - the same shape the GOD Scheduler uses. Then it applies an "ARC warp": a normalized logistic curve anchored to the endpoints (progress 0 and 1 are untouched) that gently reshapes the middle, with the amplitude scaling off the schedule's dynamic range. Finally it forces strict monotonic decrease, nudging any step that would rise back down.

The practical effect is a mid-schedule redistribution of denoising effort - a mild, deterministic cousin of the Karras-style reshaping you already know from the sampler dropdown. denoise below 1.0 compresses the whole curve toward the start, which is how img2img-style partial denoising slots in.

Inputs and output

• model - a MODEL. It's used for the sigma bounds, not actually run; you can't skip it, so wire whatever you're about to sample with.
• steps - 20 default; the number of schedule points.
• denoise - 1.0 default; lower it to shorten the travelled path.

Output is SIGMAS, wired into the sigmas input of a KSampler (Advanced) (or a custom sampler node that takes a schedule).

Installing it

Part of the Switchblade Pack: ComfyUI Manager → search MilitantHitchhiker-SwitchbladePack → install → restart, or clone manually:

cd ComfyUI/custom_nodes
git clone https://github.com/MilitantHitchhiker/MilitantHitchhiker-SwitchbladePack

Restart, then the nodes appear under MilitantAI/Switchblade/Generation. No dependencies, no downloads.

The honest take

This is where the pack's sampler corner needs a reality check. Custom sigma shapes are the one place where the KB's sampler guidance bites hard: aggressive schedule reshaping is actively harmful on flow-matching models (Flux and friends), where the community consensus is that reshaped curves like Karras break outputs - and ARC is exactly a reshaped curve. On SD 1.5 and SDXL, though, schedule tuning is a legitimate, low-stakes lever, and this is a safe place to try it because the warp is deliberately gentle (amplitude capped around 0.12–0.32).

Also be clear-eyed about provenance: there's no community track record here. A search of the usual forums turns up zero discussion of this scheduler - it's a single author's experiment shipping with a small pack. Treat it as such: A/B it against your normal scheduler at a fixed seed before you trust it, and don't be surprised if your default scheduler wins. The determinism is the feature, and it works as advertised.