H3 Soft Denoise Zone (v2v)

H3 Soft Denoise Zone (v2v) is the node you reach for when plain v2v isn't enough: instead of restyling the whole frame, it restyles one feathered region harder than the rest, and the denoise strength itself fades across the falloff. No matte line, no hard-edged cut between "changed" and "untouched". It's differential diffusion - the technique that gave images soft regional control - adapted to H3's packed audio-video latent, and per the README it's the first of its kind on H3. Marked experimental, but it's a genuinely useful one.

What it's for

Concrete use: you have footage of a person walking through a street, and you want to change just the person - swap the outfit, restyle their character - while the street stays put. Give the node a circle around them, or a per-frame mask tracking them, and the inside reinvents at the sampler's denoise while the outside drifts at a fraction of it. outer_denoise 0 makes the rest of the frame exactly the footage, untouched.

How it works

The node doesn't touch the sampler. It wraps your model: it builds a per-pixel denoise map - a circle from center_x / center_y / radius / feather (smoothstep falloff, so the strength itself fades), or from your mask - then patches the model's forward pass. A pixel only starts participating in denoising once the sampling progress threshold reaches its map value; on the input side, not-yet-participating pixels are re-noised to the current sigma (the flow forward process), and on the output side their denoised prediction is pulled back toward the clean footage. That's why the boundary is soft instead of a seam: there's no single "swap" moment, just a gradient of participation.

The inputs that matter

• v2v_latent - the Guide node's LATENT output with a v2v source active; the same latent you feed the sampler. It's the clean footage the protected regions hold on to.
• center_x / center_y / radius / feather - the circle. radius is a fraction of the short side (0.5 spans it edge to edge); feather is how much of that is falloff.
• inner_denoise / outer_denoise - fractions of the sampler's denoise. The sampler stays the master dial; inner 1.0 / outer 0.3 at sampler denoise 0.6 reinvents the centre at 0.6 while the surroundings shimmer at ~0.18.
• mask (optional) - replaces the circle. A single mask applies to the whole clip; a per-frame batch (SAM2 video segmentation) follows the subject through time. mask_feather grows and softens a hard segmentation matte outward.
• noise_seed - deterministic; change it only to reroll how the protected regions shimmer.

One output, MODEL - wire it to the sampler in place of the plain model.

Wiring it up

Guide LATENT → v2v_latent and the sampler; model → this node → sampler. Feed the same mask to H3 Regional Prompt if you also want to say what belongs in the region rather than only where change is allowed.

Where people get burned

The most common error is feeding the wrong latent - it needs H3's AV latent from the Guide node with a v2v source active, not an empty one, or you'll get a shape error. If the region isn't tracking a moving subject, that's the single-vs-batch mask distinction: one mask for a moving person means the zone stays still. And if you set inner_denoise below outer_denoise you've inverted the zone into a protector - legal, the log even tells you it's checking, but it's usually a mistake.

Install

Manager (search "H3 Studio") or cd ComfyUI/custom_nodes && git clone https://github.com/shootthesound/ComfyUI-H3Studio, restart, hard-refresh (Ctrl+Shift+R) for the frontend extension. Needs a ComfyUI build with MiniMax H3 support (v0.30.0+); no extra Python dependencies.