AI Alpha Mattes Without a Green Screen: Pulling Hair, Smoke and Glass From Ordinary Footage With LTX-2.5 Alpha-Gen
A free add-on that turns a shot you already have into a black-and-white matte, aimed at the edges keyers hate. The steps in ComfyUI, the checks before you trust it, and the one decision it will not let you make.
Every keyer ever built works by subtraction. Point it at a color, usually green, and it removes that color and keeps the rest. The whole ritual of the green-screen shoot, the lighting, the spill, the distance between subject and wall, exists to give that subtraction something clean to work with.
Take the green away and the keyer has nothing to subtract. That is why a shot of someone standing in a real street, hair lifting in the wind, has meant a day of rotoscoping for as long as there has been compositing.
Lightricks published an add-on in late September that does not subtract anything. LTX-2.5 Alpha-Gen is a video model that has been taught to look at an ordinary shot and draw a second video: white where the subject is, black where it is not, and shades of grey for everything in between. No color to key, no mask to draw, no prompt to write. The model card lists exactly the material compositors dread: hair, fur, smoke, fire, sheer fabric, glass and water.
Why this is the right kind of AI tool
Most AI video tools hand you a finished picture. You accept it or you roll the dice again.
Alpha-Gen hands you something different: a matte. That is an in-between file, the kind of thing a compositor builds a shot from, not the shot itself. You can blur it, choke it, grade it, combine it with a hand-drawn mask on the frames where it struggles, and throw away the parts you do not like. The final image is still yours.
I think that is the healthiest relationship an AI tool can have with craft work. It does the tedious first pass and leaves the judgment where it belongs.
How it works, in plain terms
LTX-2.5 is a large open video model. Alpha-Gen is a small add-on file that changes what the model is asked to do. Normally you describe a shot and it draws one. With this add-on attached, you hand it a finished clip as a reference, leave the description blank, and it draws the matching black-and-white matte, frame for frame, at the same size as your source.
Because it is a video model rather than a still-image cutout tool, it sees the whole clip at once. In principle that is what keeps the edge from shimmering frame to frame, the classic giveaway of running a photo cutout tool over video one frame at a time.
The add-on itself was created on September 28 and is version 0.9 by its filename, ltx-2.5-22b-ic-lora-alpha-gen-0.9.safetensors. Treat it as a strong preview, not a finished product.
Put this into practice
Step 1: Pick the right shot
This is the most important step, because of the one thing the add-on will not let you do: choose. The card says it plainly: "there is no prompt or mask control, and the model decides the foreground on its own."
So start with shots that have one obvious subject. A single person against a busy background. A dog on a lawn. A glass being filled on a table. Avoid two people talking, a crowd, or a shot where the "subject" is ambiguous. The model will pick something, and it may not be the thing you wanted.
Step 2: Trim and size the clip
The card sets hard limits, and both matter.
Length: 145 frames at most. At 24 frames per second that is about six seconds. The card warns that above this, "RGB content starts leaking into the matte," meaning bits of the actual picture show up where clean black and white should be.
Size: no larger than 1920 by 1088, with both width and height divisible by 32. If your footage does not fit, pad it with black bars rather than stretching it. Stretching changes the shapes the model sees, and you will have to stretch the matte back afterwards.
For a longer shot, my suggestion (not the card's) is to cut it into overlapping six-second pieces and blend the mattes where they overlap. Expect to check the joins by eye, since the model has no idea the pieces belong together.
Step 3: Run it in ComfyUI
The setup follows the card:
- Put the add-on file into ComfyUI's
models/lorasfolder. - Load the fast version of LTX-2.5 (listed as the 22B "distilled" model).
- Open the
LTX-2.5_V2V_ICLoRA_Single_Stage_Distilled.jsonworkflow from Lightricks' ComfyUI-LTXVideo project. - Connect your clip as the reference video, with its strength at 1.0.
- Set the add-on's strength to 1.0 and leave the prompt completely empty.
- Run only the first stage, at your clip's native size. The card is specific that the second, upscaling stage is skipped.
What comes out is a grayscale video the same length as your source, at the size you fed in. If you padded the clip, crop the matte by the same amount afterwards.
Step 4: Check it before you trust it
Never judge a matte against the original picture. Your eye fills in what it expects. Instead:
- Use the matte to place your subject over flat black, then flat white, then a saturated red. Each background exposes different failures: dark fringes show on white, light halos show on black, and red is unforgiving with hair.
- Scrub frame by frame through the hardest moments, usually where hair crosses a bright background or smoke thins out. Watch whether the edge holds steady or crawls.
- Look at the end of the clip especially. If anything is going to start leaking, it will show up as the frame count climbs.
Step 5: Use it in your compositor
In Resolve, After Effects, Nuke or Fusion, bring the matte in as a second clip and use it as a luminance matte for the original footage (a matte that reads brightness as see-through-ness), so white keeps and black removes. From there it is ordinary compositing: soften the edge slightly, pull the matte in a pixel or two if there is a halo, and paint in fixes on the handful of frames that need them.
That last part matters. The goal is not a perfect matte from the machine. The goal is a matte that is 90% of the way there, so your afternoon goes into the hard 10%.
When you should still shoot green
None of this retires the green screen, and I would be wary of anyone who says it does.
If you control the shoot, a well-lit green screen still gives you something Alpha-Gen cannot: certainty. A keyer does exactly the same thing on frame one and frame four thousand, on every take, on every machine. You can predict its failures before the shoot and light around them. An AI matte is a guess, a very good one on the right material, and a guess you have to check frame by frame.
Where Alpha-Gen earns its place is everywhere a green screen was never an option. Documentary footage. Run-and-gun B-roll. Archive material. The client's phone video. The beautiful take where the actor stepped off the green and into the real room. The plate from a location where a screen would have taken an hour to rig and the light was going.
That is a lot of footage. For a lot of working editors and motion designers, it is a big share of what lands on their desk. A tool that gives those shots a fighting chance is worth learning, even if the studio shoot keeps its green.
Where it breaks
It chooses the subject, not you. Covered above, and it is the limit that will bite most often. There is no way to say "the woman on the left."
Six seconds is the safe ceiling. Beyond 145 frames the matte gets contaminated. Long takes mean splitting and blending, which is your work.
Full HD needs serious hardware. The card states that full-HD clips at the full 145 frames "need an H100/B200-class GPU," the kind of card that lives in data centers. Most people will rent one by the hour or run at lower resolution. The card does not publish memory figures for smaller sizes, so test on a short clip before you plan a job around your own card.
There are no accuracy numbers. Lightricks shows examples (backlit hair, a dancer in smoke, dog fur, fire, a sheer curtain, pouring water) but no measured comparison against other matting tools. Your own footage is the only test that counts.
It is a 0.9. The filename says so. Expect a revision, and keep the version you tested with if a job depends on it.
The license has one line to know. It falls under the LTX-2 community license, which says Lightricks "claims no rights in the Output you generate." Businesses with annual revenue of $10 million or more need a paid license. For freelancers and most small studios, the mattes are yours to deliver.
The habit worth building
Rotoscoping hair has always been the job that tells you whether a shot was worth fixing. Tools like Alpha-Gen move that line. A shot you would have reshot or abandoned last year might now be an hour of cleanup.
So here is a small test. Find the worst shot in your archive, the one with flyaway hair against a bright window that you never managed to key. Trim it to six seconds, run it through Alpha-Gen at whatever resolution your machine or a rented card can handle, and put the result over red. Whatever you see, you will know where AI matting really stands on your own footage, which is worth more than any demo reel.
Medium metadata
Title: AI Alpha Mattes Without a Green Screen: Pulling Hair, Smoke and Glass From Ordinary Footage With LTX-2.5 Alpha-Gen
Subtitle: A free add-on that turns a shot you already have into a black-and-white matte, aimed at the edges keyers hate. The steps in ComfyUI, the checks before you trust it, and the one decision it will not let you make.
Tags: Compositing, Visual Effects, Video Editing, AI Tools, Filmmaking
Estimated read time: 8 minutes