Comfy Pro Alpha Generator

If you came here looking for SelectiveMatte, the first useful thing I can tell you is that the node changed its name. SelectiveMatte is the class id this thing shipped with, and it still works - ComfyUI keeps it registered so workflows saved months ago open and run. In the node menu you'll find it as Comfy Pro Alpha Generator, under the Comfy Compositing category. Same node, same inputs, new id.

What it does is the annoying half of compositing: pull one chosen person out of a plate, with hair. Not "the foreground" - one subject, of your choosing, with generative-quality edges.

Why this node exists at all

Background removal is the most commoditised job in this ecosystem, and the frontier moved years ago from "can it find the subject" to "can it survive hair, fur and semi-transparency" (that framing is straight out of our background removal writeup, and it's still accurate). SAM3 solved the finding. It did not solve the edges - a mask is a per-frame measurement, and measurement doesn't invent strands that the segmentation outline never contained.

Meanwhile the generative approach got the edges right and the subject wrong: LTX-2.5's alpha-gen IC-LoRA renders hair-level alpha, but it mattes every main subject in the frame and there's no way to point at one. Three kids with their arms around each other? You get three kids.

This node is the splice. SAM3 decides which subject. LTX decides how the edge looks. The author measured the gap on a 247-frame plate: a plain SAM3 mask recovers 29% of the hair outside its own outline, SAM3 + ViTMatte gets to 45%, and this gets to roughly 89%.

How the mechanism actually works

The trick isn't "run both and blend". It's that LTX can't be told which subject to matte, so the node changes what LTX sees. SAM3 finds every other subject and they get painted out of a copy of the plate with inpainted background - not a flat colour fill, because LTX reads a flat fill as a semi-transparent object and mattes it wrong. LTX then mattes your subject alone, and draws all of its edges, including the parts SAM3 missed.

Where your subject touches someone else, LTX returns 1.0 on both sides and there's no edge to take, so within a small band the node substitutes a stabilised ViTMatte edge. A low-res coverage map catches the pieces LTX refused to read as a main subject - a parked car, a table - and fills only those with ViTMatte. SAM3 and ViTMatte work frame by frame and therefore chatter, so their output is warped and averaged across neighbouring frames with optical flow before the blend. LTX renders 49 frames in one go, so its edges are already steady.

At run time the node expands into a full graph of SAM3, ViTMatte and LTX nodes, each cached separately. That's why flipping quality after a render is cheap: the passes that didn't change get reused.

The inputs you'll actually touch

images is the plate - Load Video → Get Video Components, a Load Plate Sequence, or any crop/resize chain. fps and first_frame ride along from whichever loader you used; first_frame is what keeps VFX numbering (1001 in, alpha_01001 out).

You don't set picks by hand. The viewer writes that JSON when you click the subject - click to pick, shift-click to add a part SAM3 missed, alt-click to cut someone else's hand out of the mask.

quality is Best (v12, 2 LTX passes) or Fast (v7, 1 LTX pass). Fast is roughly half the time and fine for look-dev; edges against other people are a bit weaker. render at preview chunk gives you the first 49 frames from your pick frame and appends _preview to filename_prefix.

The model selectors expect specific files (ltx_model, alpha_lora, video_vae, audio_vae, vitmatte) and show MISSING: <file> rather than silently falling back to something unrelated - which matters, because a Flux VAE crashes deep inside LTX with a useless error. text_encoder defaults to the built-in empty prompt; leave it there, alpha-gen's prompt is empty and the pack ships that conditioning so nobody loads a 13 GB encoder for nothing.

Under the optional group sit the interesting switches: overlap_mode for people crossing your subject, native_edges for plates above 1080p, edge_colour (quick / perfect / off), which rebuilds the subject's own colour in the soft edges, clean_plate_lora for perfect, plus the tuning that only matters once overlap mode is on - v13_depth, hair_guide, attn_grow, occluder_reach, limb_cap.

Outputs

Two: alpha (MASK) and comp (IMAGE). alpha is the 16-bit matte, and the node also writes files to ComfyUI/output/<filename_prefix>/ itself - png16 or half-float exr, plus rgba_ comp frames since v0.5. The comp output is only built if you actually wire something to it; on a full 1080p shot that's several gigabytes, so the node checks first.

Install

ComfyUI Manager → Install via Git URL → https://github.com/rik-python/ComfyUI-Comfy-Pro-Alpha-Generator.git. 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 install ComfyUI-SAM3 and ComfyUI-LTXVideo, accept the two gated Lightricks licenses on Hugging Face, and download about 25 GB of weights into models/diffusion_models, models/loras and models/vae. SAM3, ViTMatte and Depth Anything download themselves. You need ComfyUI 0.38 or newer - the node is built on the V3 node API. The pack is free for personal and non-commercial use; commercial work needs a license from the author.

Where people get burned

The one that wastes an afternoon: you crop the plate (or swap the clip) after picking, and the picks still look fine but land in the wrong place. The node fingerprints the plate and prints the plate changed since the subject was picked to the console - check there before you blame the model.

After a ComfyUI restart, the small proxy images the picker scrubs are gone from temp/, so the viewer can only show boxes until you click Load plate again. Normal, not a bug.

Long shots hit a hard ceiling: more than 24 chunks and the node stops with a message telling you to lower overlap frames or split the shot. At the defaults that's about 960 frames.

And if alpha is empty for the first part of your shot - it's meant to be. The matte starts at the frame you picked on; everything before it comes out empty.