Comfy Pro Alpha Generator

Every matte tool in ComfyUI makes you pick a side. A segmentation model - SAM 3, BiRefNet, whatever's winning this month - finds exactly the person you clicked and then hands you a soft outline that eats the hair. A generative matte draws hair down to the individual strand and then refuses to tell you which of the three kids in the frame it matted. Comfy Pro Alpha Generator is a pack by Rikin Patel (Comfy Compositing, and yes, he's a Nuke compositor) built to stop choosing: SAM 3 decides who, LTX-2.5's alpha-gen IC-LoRA decides what the edge looks like.

The gap it's aiming at is real. Vibe-masking got commoditised years ago; the frontier moved from "can it find the subject" to "can it survive hair, fur and semi-transparency", and nobody won that. This node's pitch is that hair is a rendering problem, not a segmentation problem.

The node's own headline numbers, measured on a 247-frame plate where only the middle kid should be matted: 88.7% edge-detail match against pure LTX alpha-gen (the reference for hair), versus 45% for the SAM 3 + ViTMatte route and 29% for a raw SAM 3 mask. Flicker lands at 1.2x LTX alone. Salt, as usual: it's the author's own harness and it favours LTX-shaped edges. The visual difference between strands and halo isn't subtle either way.

How it actually gets there

On Queue, the node expands into a full graph of SAM 3, ViTMatte and LTX stages. You never wire that by hand - you get one node with a picker and a render button.

The core trick is that LTX can't be told which subject to matte, so the node changes what LTX sees. SAM 3 tracks every other subject in the plate, and those get painted out of a copy with an inpainted background - a flat colour fill doesn't work, LTX reads it as a half-transparent object. LTX then mattes your subject alone and draws all its edges, including the parts SAM 3 missed. Where the subject overlaps another person, LTX outputs 1.0 on both sides so there's no edge to take, and within about 15 px of other people the node falls back to a stabilised ViTMatte edge. A coarse coverage map catches the parts LTX left empty entirely - a parked car, a table - and fills only those.

LTX renders 49 frames in one go, so its edges hold still; SAM 3 and ViTMatte decide per frame, so their output is smoothed with optical-flow-warped neighbours before the blend. Long shots render in overlapping chunks that get cross-faded. Every stage is cached, which is why switching Fast/Best or running the full shot after a preview doesn't redo finished work.

The inputs you actually touch

images is the plate - Load Video → Get Video Components, or Load Plate Sequence, or any crop/resize chain. fps and first_frame come along with it; first_frame is the one that makes VFX numbering work (1001 gives you alpha_01001).

quality picks 2 LTX passes (Best) or 1 (Fast, roughly half the time, weaker edges against other people). render chooses a 49-frame preview from the pick frame or the full shot. picks is the JSON the picker writes when you click - don't hand-edit it. filename_prefix puts each shot in its own output folder, and format is png16 or exr.

Then a row of model selectors whose only real job is telling you a file is missing: ltx_model, alpha_lora, video_vae (must be the conv VAE), audio_vae, vitmatte, and text_encoder, which you leave on built-in empty prompt forever because alpha-gen runs with an empty prompt and the pack ships that conditioning. Under the Advanced fold: chunk, overlap, seed, erode/dilate, near_px, stabilize_radius, sam3_threshold, fast_attention. Optional switches handle the specialist cases - overlap_mode for crossing people, native_edges for plates over 1080p, edge_colour for the soft-edge halos.

Outputs

alpha is the 16-bit MASK. comp is the subject composited over grey - and it's only computed if something is wired to it, because a full 1080p shot of frames is several gigabytes. Leave it disconnected unless you're previewing.

Install

Manager → Install via Git URL → https://github.com/rik-python/ComfyUI-Comfy-Pro-Alpha-Generator.git, then install ComfyUI-SAM3 and ComfyUI-LTXVideo the same way. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/rik-python/ComfyUI-Comfy-Pro-Alpha-Generator.git
pip install -r ComfyUI-Comfy-Pro-Alpha-Generator/requirements.txt

Then the models, about 28 GB: the LTX-2.5 int8 transformer into diffusion_models/, the alpha-gen IC-LoRA into loras/, the conv video VAE and audio VAE into vae/. SAM 3, ViTMatte and Depth Anything V2 Small download themselves on first run. Both Lightricks repos are gated - you need a free Hugging Face account, the terms accepted on each page, and hf auth login if you're pulling from the CLI.

Where people get burned

MISSING: <file> in a model selector stops the run with the download link instead of failing halfway, so read it rather than guessing. Frames before the frame you picked come out empty - tracking only runs forward. A shot needs more than 24 chunks (roughly 960 frames at defaults) and it hard-errors; split it. 4K with native edges holds every frame in RAM, 97 GB at 242 frames, so split long 4K shots into ~100-frame runs.

And the licence catches people off guard: the pack is PolyForm Noncommercial - paid client work needs a commercial licence from the author - and LTX-2.x adds its own terms on top, including a paid tier above $10M annual revenue. Check both before you ship.