H3 Transparent Video - SatoDive

You cannot get an alpha channel out of a diffusion video model. Video transparency has been its own research line for years - TransPixar and friends - and the practical route everyone uses is the boring one: generate over a flat backdrop, then key it. H3 Transparent Video is that second half, in one node.

Why it matters. The instinct is to reach for a segmentation model instead - BiRefNet, InSPyReNet, RMBG - and it fails on video for a structural reason: a segmentation model labels each pixel foreground or background, and matting predicts fractional alpha. Smoke, motion blur, a thin veil, a spark trail are neither solid nor absent, and a hard mask can't say so. Keying against a modelled backdrop can.

How it works

A generated clip never has a truly flat background - it has gradients, vignettes and shading - so the keyer models the backdrop instead of assuming a colour. keyer.py reads the frame border across up to a dozen frames, robustly, and reports a border-match confidence. It then estimates a per-frame clean plate at low resolution from confident backdrop pixels, fills the holes the subject punched in it with push-pull, and smooths it in space and time. Every pixel is compared against its own local backdrop. That is the mechanic that kills haze: a flat key colour leaves a faint veil wherever the generation was slightly off-green; a per-pixel plate doesn't.

The key itself is colour-difference "screen" keying for green and blue, unmultiply for glow on black, and the mirror for smoke and ink on white. Then the cleanup: a colour-guided filter that snaps the matte to full-resolution edges (repairing 4:2:0 chroma blockiness), motion-gated temporal smoothing so moving edges don't ghost, despill, a proper un-mix of the foreground out of the plate - F = (C - (1-a)P) / a - and colour extension under near-transparent pixels so edges don't pick up a dark fringe when you scale or composite. The docstring makes an unusually honest argument for one pass over two-render triangulation: a diffusion model can't re-render a background behind a held subject, because holding the subject holds whatever was behind its soft edges.

The inputs you'll actually touch

backdrop first, matching what H3 Shot wrote into your prompt; Auto reads it from the border and usually gets it right. Then images (frames straight out of VAE Decode - no intermediate MP4 loss) or video, plus audio. output_format is the real decision: ProRes 4444 for Premiere/AE/Resolve/Final Cut, RGBA PNG sequence, or "Don't save (preview only)". filename names a numbered folder under output/H3_Transparent_Video/; fps at 0 takes the rate from the video input, or assumes 24 for images; leave alpha_type on Straight.

Everything else you touch when something is wrong. screen_gain is the master strength - raise it if a faint veil survives, lower it if thin or soft parts vanish. plate_fix at 1 is the full per-pixel backdrop model. clip_black (0.03) clears leftover haze; clip_white (0.95) fills faint holes. despeckle, despill, choke, softness, shadows, temporal and edge_solve are the usual keyer controls, and keep_mask / garbage_mask accept masks from any segmentation node.

Outputs

foreground (RGB), alpha (MASK), rgba (composite-ready frames), preview (composited over what preview_on says - Checkerboard, Grey, Black, White, or your preview_plate), and report. Read the report: backdrop kind and hex, border-match percentage, haze and flicker numbers, plate unevenness, and any warnings. It also lands in report.txt next to the renders.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/SatoDive/Minimax-H3-Transparent-Video

Restart. No extra Python packages - torch, numpy, Pillow and av, all bundled - and no model downloads. The README points at a ready workflow (workflows/H3_Generate_Transparent_Clip.json, plus a Patreon download); the chain it documents is H3 Shot → sampler → VAEDecode → this node, with VAEDecodeAudio alongside.

Where it bites

Leftover dots after keying is the classic complaint, and not specific to this pack - a July 2026 thread about Krea 2 green-screen output had the same symptom in Resolve, with commenters split between matting nodes and "green screen is honestly more hassle than just masking directly in post." Cheap fixes first: raise screen_gain, lower clip_black, raise despeckle, and read the report's border-match percentage.

Holes in the subject mean it contains the backdrop colour - green clothing on green screen. Switch to Blue, use keep_mask, lower clip_white, or fill_holes.

Fire, sparks and orange glow belong on Black, not green; dark-grey objects on White come out partly transparent by design, because that mode is for ink, smoke and dust.

Check the report if a Pro-only setting looks ignored. The source has an edition switch: on a Free build, custom colour falls back to Auto, edge solve / temporal / fill holes / keep and garbage masks are skipped, WebM and QuickTime Animation save as ProRes 4444, and only the first 124 frames get keyed.