Radiance Shift Scheduler

Radiance Shift Scheduler is the "trust me, I tuned it" version of the same pack's Power Shift Scheduler. Same double power curve, same SIGMAS output, same family. But where Power Shift ships with neutral defaults (power 2.0, midpoint 1.0, discard_penultimate off), Radiance ships with the author's preferred settings baked in: power 2.4, midpoint_shift 0.98, and discard_penultimate on. If you want the power-shift look on a Flux-style model without dialing knobs, this is the one to drop in first.

What's actually different

Under the hood it's the same ts = (1 - x**power)**power reshaped-sigma trick, with two tweaks you can feel. It computes one extra internal step and then trims back to your steps, so the curve isn't quite the same shape as Power Shift at equal step counts - the rounding lands differently. And because discard_penultimate defaults to true, it skips the last near-zero denoising step by default, which trims a hair off the end where the sampler is mostly refining nothing.

The result: a schedule that spends proportionally more effort at the extremes and less in the mushy middle - the shape people reach for when a flow-matching model like Flux or Wan comes out washed out or lacking punch at low step counts. Think of it as the author's opinionated preset. Leave everything default, see how it looks, and then only touch the two knobs that matter:

• power (default 2.4): curve sharpness. Up for a more extreme S-curve, down toward 1.5 for something gentler.
• midpoint_shift (default 0.98): skew. Nudge below 1 to throw more steps at the detail end.

The rest of the inputs match its sibling - model (the checkpoint you're sampling, which provides the sigma ladder it reshapes), steps, discard_penultimate, and denoise for img2img/hires passes. Output is a single SIGMAS tensor for SamplerCustomAdvanced.

Install

It ships in the same pack, so one install covers all four nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI_PowerShiftScheduler

or search "ComfyUI Power Shift Scheduler" in ComfyUI Manager, then restart. No model downloads. The pack imports scipy but ships no requirements.txt, so if the nodes never appear, that's your missing dependency - pip install scipy in the ComfyUI environment.

Gotchas

Two things trip people up. First, this is a scheduler, not a sampler - it shapes when denoising happens, so pair it with whatever sampler your model card already wants (Euler-family on flow models, per the usual guidance). Second, installing the pack also adds radiance_shift as a name in the plain KSampler scheduler dropdown, so you can try it without touching the custom sampling stack at all - useful as a quick A/B against your current scheduler before you commit to wiring the node in.