Sigmas LinearSine

This is the most experimental of RES4LYF's little sigma nodes, and it wears it on its sleeve. Sigmas LinearSine overlays a sine wave onto your noise schedule - so instead of a smooth monotonic descent, the sampler walks down a curve with a controllable ripple in it. It blends that oscillation with a plain linear ramp so it doesn't go completely off the rails. This is squarely in "modulate parameters vs. time" territory, which the author calls out as a deliberate focus of the whole pack.

Why would you want a wavy noise schedule? Honestly, for most workflows you don't - the flow-matching wisdom is that these models like balanced schedules and dislike aggressive reshaping. But the pack is built for people who want to push exactly that boundary, and periodic modulation of the denoising effort can, on the right model and prompt, shake loose texture and variation that a flat descent won't. Treat it as a toy for exploration, not a default.

How it works

You give it a SIGMAS schedule, and the node computes a sine wave across the steps and mixes it into the curve. The linear_weight sets how much plain linear schedule survives versus how much sine you're injecting - so at one extreme you get essentially your original ramp, and at the other you get a much more oscillatory path. Amplitude, frequency, and phase shape the wave itself.

The inputs that matter

Five inputs, and you'll mostly play with two or three:

• sigmas (SIGMAS) - the schedule to modulate.
• amplitude (default 0.5) - how tall the ripple is. This is the strength dial; keep it low unless you want obvious effects.
• frequency (default 1) - how many oscillations across the schedule.
• phase (default 0, roughly -2π to 2π) - where in the wave you start. Usually leave it.
• linear_weight (default 0.5, range 0–1) - the blend between a straight linear ramp and the sine. This is the safety valve: push it toward 1 to keep things sane, toward 0 to let the wave dominate.

Output is one SIGMAS, wired into your sampler's sigmas input, which overrides the scheduler and steps set inside the sampler.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/ClownsharkBatwing/RES4LYF/, activate the venv, pip install -r requirements.txt (portable build uses the embedded python's pip), restart, hard-refresh with F5. For the nested sampler menu, add rgthree-comfy with "Auto Nest Subdirectories" on.

Common issues

The big one: too much amplitude or too little linear weight and the schedule stops being monotonic - it can rise where it should keep falling. Hand a sampler a schedule that goes up mid-run and you'll get artifacts, mush, or plain garbage, because the model is being told to add detail back it just removed. If your output falls apart, dial amplitude down and linear_weight up until it behaves, then creep back out.

And set expectations. This is not a "make my images better" node; it's a "let me see what happens if the noise curve breathes" node. Keep a fixed seed while you sweep the parameters, or the sine effect and seed-to-seed variance will blur together and you'll learn nothing. If you just want solid results on a flow-matching model, a clean scheduler like beta57 or simple will serve you far better than this ever will.