Nodes/ComfyUI-H3Studio/H3 Soft Denoise Zone (v2v)
ComfyUI Node

H3 Soft Denoise Zone (v2v)

Spatially-varying denoise for H3 v2v: a feathered zone (circle or mask) restyles harder than the rest of the frame, with the strength itself fading across the feather — no matte lines. Differential diffusion, adapted to H3's packed AV latent.

By shootthesound·Created 13 days ago·Updated 5 days ago· 20
H3 Soft Denoise Zone (v2v)
  • model
  • v2v_latent
  • mask
  • MODEL
center_x0.50
center_y0.50
radius0.35
feather0.50
inner_denoise1.00
outer_denoise0.30
noise_seed77
mask_feather0.10
Categorymodel/conditioning/minimax

Inputs (11)

NameTypeDefaultDescription
modelMODEL
v2v_latentLATENTThe Guide node's LATENT output with a v2v source active — the SAME latent you feed the sampler. It is the clean footage the protected regions hold on to.
center_xFLOAT0.500–1Zone centre, fraction of the frame width.
center_yFLOAT0.500–1Zone centre, fraction of the frame height.
radiusFLOAT0.350.05–1.5Zone radius as a fraction of the frame's SHORT side (0.5 spans it edge to edge).
featherFLOAT0.500–1How much of the radius is the soft falloff. 0 = hard edge (matte-line territory), 0.5 = the outer half fades, 1 = fades from the centre out.
inner_denoiseFLOAT1.000–1Fraction of the SAMPLER'S denoise applied inside the zone. 1.0 = the full wired denoise; the sampler stays the master dial.
outer_denoiseFLOAT0.300–1Fraction of the sampler's denoise outside the zone. 0 = the surroundings reproduce the footage untouched.
noise_seedINT770–2147483647Seed for the re-injection noise field (deterministic; change it only to reroll how protected regions shimmer).
mask_featherFLOAT0.100–0.5Softens a MASK input (radius as a fraction of the frame's short side): the mask is grown then blurred, so a hard segmentation matte (SAM2 etc.) fades outward instead of leaving a matte line. 0 = use the mask as-is. Ignored for the circle (it has its own feather).
maskoptMASKOptional: replaces the circle. White areas get inner_denoise, black get outer_denoise, greys blend. A SINGLE mask applies to the whole clip; a mask BATCH (one per frame at 24 fps — e.g. SAM2 video segmentation tracking a person) follows the subject through time, pooled onto the latent grid. Frame counts that don't match the clip are resampled by index.

Outputs (1)

NameTypeDescription
MODELMODELWire to the sampler in place of the plain model.