Power Ease Scheduler

PowerEaseScheduler is the simpler sibling of the pack's Dual Ease Cosine Scheduler, and "simpler" is doing real work here. Where the dual version gives you two knobs for independent control at each end of the sigma curve, this one gives you a single power slider and a clean, predictable way to tilt the whole schedule. It outputs a SIGMAS tensor that plugs into any sampler that accepts custom sigmas.

The formula is about as transparent as scheduling gets: sigma = sigma_max - (sigma_max - sigma_min) * t^power, where t runs from 0 to 1 across your steps. At t=0 you start at sigma_max (full noise), at t=1 you end at sigma_min (near-clean), and the power value decides how the in-between is distributed. power higher than the default 3 pushes more of your steps toward the high-noise end of the journey; power below 1 (the floor is 0.1) crowds them toward the low-noise refinement end. One slider, one intuition: which half of the image-making do you want more time spent on?

Inputs are steps (default 20), sigma_max (14.614642) and sigma_min (0.02917), and power (default 3). There's also an optional model input with the same behavior as the dual version: connect a MODEL and it overrides your sigma_max/sigma_min with the model's own sampling range, so the sliders go quiet. If you're setting those by hand, leave the model unplugged.

Like its sibling, the output is a strictly-decreasing schedule with a zero appended Karras-style, which is the shape samplers in custom-sampling blocks expect. It lives in the pack's sigma category alongside a Sigma Visualizer that can render the schedule as a chart - genuinely useful when you're trying to understand what a power of 4 actually does to your curve before you burn a render on it.

Same caveats as the dual version, honestly. This is a tinkerer's scheduler for classic DDPM-style models - on SD 1.5 and SDXL you can push sampling effort around and see real results, but the current wisdom on flow-matching models (Flux and friends) is that heavily reshaped sigma curves usually cost you more than they buy. And the effect of power is gradual, so change it in small steps and judge on real outputs, not on the assumption that a big number equals a big win.

Install via ComfyUI Manager (search "WtlNodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Scorpiosis0/ComfyUI-WtlNodes.git

then restart ComfyUI. Deps: numpy, scipy, pillow. No models, no keys.

If you've ever wondered what a "Karras-shaped" schedule is doing to your image, this is the node that lets you take the wheel - one knob, no black box.