Nodes/ComfyUI-CoachBate/CoachBate H3 Per-Subject Denoise
ComfyUI Node

CoachBate H3 Per-Subject Denoise

Per-frame denoise strength, scaled inversely to subject size.

By CoachBate·Created 2 months ago·Updated a day ago· 0
CoachBate H3 Per-Subject Denoise
  • model
  • av_latent
  • transform
  • av_latent
  • report
  • model
◄denoise_multiplier_small_subject1.00►
◄denoise_multiplier_large_subject1.00►
◄scale_modeabsolute_px►
◄subject_px_small30►
◄subject_px_large120►
◄gamma1.0►
◄smooth_frames9►
CategoryCoachBate/H3 Subject Refine

Inputs (10)

NameTypeDefaultDescription
modelMODELThe H3 model, on its way to the sampler. Required: making a per-frame mask behave takes two changes to the MODEL rather than to the latent. This node's video mask is kept out of H3's per-token timesteps, and the frames that mask holds back are re-noised to the step the sampler is on instead of being handed back nearly clean. Pass the returned model on to the guider.
av_latentLATENT—
transformH3FACEXFORM—
denoise_multiplier_small_subjectFLOAT1.000–1MULTIPLIER on the denoise set on BasicScheduler, applied where the subject is SMALLEST. Not a denoise value in itself: at 1.0 those frames get the full BasicScheduler denoise, at 0.8 they get 80% of it. Below 1.0 NO frame in the clip receives the FULL denoise you set, which makes the whole clip gentler.
denoise_multiplier_large_subjectFLOAT1.000–1MULTIPLIER on the denoise set on BasicScheduler, applied where the subject is LARGEST. Lower preserves the real detail those frames already have, since a big subject needs less rebuilding than a distant one.
scale_modeCOMBOabsolute_pxabsolute_px: strength is set by real subject size in SOURCE pixels via subject_px_small/large. Safe across a batch - a video that never has a small subject gets the baseline throughout. relative_to_clip: normalise to this video's own min/max, so its smallest subject always gets the full boost regardless of actual size. Use when tuning a single video to its extremes.
subject_px_smallFLOAT304–400Subject height (SOURCE px) at or below which the full denoise_multiplier_small_subject is applied. Genuinely tiny subjects only.
subject_px_largeFLOAT1208–800Subject height (SOURCE px) at or above which denoise_multiplier_large_subject is applied. Calibrated so a video whose smallest subject is ~90px gets only a mild boost - that size was already fine at the baseline denoise - and anything past 120px gets none.
gammaFLOAT1.00.2–4Curve on the interpolation. >1 keeps strength high until the subject is genuinely large; <1 drops it off early.
smooth_framesINT91–61Smooth the strength curve over time. An abrupt change in denoise between neighbouring frames is visible as a texture pop, so this wants to be generous.

Outputs (3)

NameTypeDescription
av_latentLATENT—
reportSTRING—
modelMODEL—