capacap 3d4283dd43 Taper schedule tail and log per-run pass stats
Linearly decays intensity across the last ~30% of steps and zeroes the
final ~15%, so those steps skip the second model call via the existing
adjustment==0 fast path. Saves 14-17% wallclock on typical workloads and
softens the over-sharpening that showed up at higher step counts.

Also adds a per-run stats line logging detail vs. skip counts so the
impact of future schedule/threshold tweaks is visible. Includes a
mask-activity floor (currently 0.02) as a secondary skip path; not
tripping in practice but kept as a safety net for degenerate masks.
2026-04-21 18:13:03 +02:00

ComfyUI-Selective-Sigma-Detailer

A ComfyUI custom sampler that boosts detail only in latent regions that are already dense at the moment its schedule activates. Smooth regions (skies, out-of-focus backgrounds, flat surfaces) are left untouched; busy regions (faces, fabric, foliage) get a sigma-shifted pass that sharpens them.

How it works

  1. Early sampling steps run unmodified so the model can establish composition.
  2. On the first step where the schedule becomes active, a density mask is derived from the model's denoised prediction via local latent variance, and then frozen for the rest of the run.
  3. For every subsequent active step, two denoises are computed (normal, and one at a shifted sigma) and blended per-pixel by the mask.

Cost: 2x model calls on active steps only. Inactive steps are free.

Node

Selective Sigma Detailer (sampling/custom_sampling/samplers)

Wraps a SAMPLER and returns a new SAMPLER. Drop it between your base sampler (KSamplerSelect, etc.) and SamplerCustom.

Schedule parameters

detail_amount, start, end, bias, exponent, start_offset, end_offset, fade, smooth, cfg_scale_override control when and how strongly the detail pass fires, following the same schedule shape as Detail Daemon.

Mask parameters

  • variance_kernel — window size for local variance (odd, 3–15). Larger is coarser.
  • mask_blur — smoothing applied after variance. Softens transitions.
  • mask_threshold — values below this are pulled to 0. Higher = stricter targeting.
  • mask_gamma — contrast curve on the mask. >1 sharpens, <1 softens.
  • save_mask_preview — dumps the frozen mask as a PNG to ComfyUI's temp dir.

Credits

Inspired by ComfyUI-Detail-Daemon by Jonseed; the schedule shape used to gate the active window follows the same idea.

License

MIT.

S
Description
No description provided
Readme MIT
12 MiB
Languages
Python 100%