Sigmas Iteration Polyexp

Like Sigmas Iteration Karras, this node builds a noise schedule that goes up and then down instead of the usual one-way "full noise to clean image" curve - a Karras-shaped ramp exists as a sibling node with the same up/down structure, and this is its poly-exponential counterpart, swapping in a different curve family for each half of the round trip. That up-then-down shape is the backbone of RES4LYF's unsampling and resampling workflows, where you deliberately re-noise an existing image and re-solve it to edit or refine it - the README's own framing is a pile of "sigma, latent, and noise manipulation nodes" built specifically because sculpting parameters against time is where a lot of the pack's quality gains come from.

How it works

It generates two segments and stitches them: steps_up steps climbing toward s_max, then steps_down steps falling back down. Each half is shaped by its own rho parameter (rho_up and rho_down), the same curvature-control role rho plays in a standard Karras schedule - how aggressively the change concentrates toward one end of the ramp rather than spreading evenly. The "polyexp" name points at a polynomial-plus-exponential curve construction rather than the pure power-law Karras uses, which in practice means a different feel to how the noise ramps in and back out, even though the parameter shapes (steps up/down, rho per side, min/max bounds) mirror its Karras sibling closely. Alongside the sigmas, it also produces a parallel momentums channel for the pack's momentum-aware sampling, and can accept existing sigmas/momentums to build on rather than starting fresh.

The inputs and outputs that matter

• steps_up / steps_down (both default 30) - how many steps to spend re-noising versus denoising back down. They don't have to match.
• rho_up (default 0.6) / rho_down (default 0.8) - curvature of each half. Note these defaults are noticeably lower than Karras's typical rho values (often 3-7) - expect a visibly different ramp shape between the two nodes even at otherwise matching settings, since "rho" means something specific to each curve family rather than being directly interchangeable.
• s_max (default 2) - the peak sigma the up-ramp climbs to, i.e. how much noise gets injected.
• s_min_start / s_min_end (both default 0.0291675) - the floor sigma at each end of the round trip. That default is the same figure used as the floor across the rest of the pack's sigma-cleanup nodes, which reads like a real model's minimum sigma rather than an arbitrary placeholder.

Outputs are sigmas (SIGMAS) - the full up-down schedule for a sampler's sigmas input - and momentums (SIGMAS), for momentum-aware sampling. The optional sigmas/momentums inputs let you chain from another schedule node instead of generating fresh.

How to install it

• ComfyUI Manager - search "RES4LYF", install, restart.
• Manual - activate your venv, cd ComfyUI/custom_nodes && git clone https://github.com/ClownsharkBatwing/RES4LYF, cd RES4LYF, pip install -r requirements.txt (portable builds: use the embedded pip.exe). Restart.

Common issues & troubleshooting

This isn't for a plain txt2img sampler. Plug it into a normal, one-way sampler and the "noise back up" half of the schedule will just look broken. It belongs with sampler_mode set to unsample/resample on a RES4LYF sampler, which the pack is explicit is essential - it disables ComfyUI's own automatic noise addition so the schedule you built is the only noise source in play.

s_max is your intensity dial. Push it too high and you re-noise the image hard enough to lose most of what made the original distinct; too low and the round trip barely changes anything. There's no documented "right" value - start near the default and move deliberately, checking output at each step.

Whether to reach for this node or Sigmas Iteration Karras isn't spelled out anywhere. Both build the same up-down shape; the underlying curve math differs. If you're not sure which suits your case, that's consistent with the pack's broader documentation gap - the community's standing request for a RES4LYF scheduler guide covers exactly this kind of "which one, and why" question. Try both on the same seed and compare.