Pulling a clean subject out of a background shot without a green screen has been a compositing fantasy for decades. AlphaGen, an open-weights add-on for Lightricks’ LTX-2.5 video model, takes a real swing at it: feed it an ordinary RGB clip and it predicts a grayscale alpha matte — white where the model reads foreground, black where it reads background, gray for partial transparency at edges like hair or motion blur. No mask, no prompt, no key color. This week it landed as an official part of ComfyUI’s LTX VFX node bundle, alongside Reframe, SDR-to-HDR, and Upscale — four finishing tools comfy.org frames as one coherent toolkit for a post pipeline.
What it actually does, and what it doesn’t
AlphaGen is an in-context LoRA (IC-LoRA) trained on top of the 22-billion-parameter LTX-2.5 base, not a standalone network. According to Lightricks’ own technical documentation, the workflow runs with an empty prompt and the adapter strength fixed at 1.0 — the source video is the only guide the model gets. That’s a meaningfully different shape from a typical generative tool: you’re not describing what to keep, the model is deciding, which is also the honest limit of the approach. Comfy.org’s own framing promises “full-resolution output up to 4K,” but Lightricks’ documentation states the model itself caps inputs at 1920×1088, padded to dimensions divisible by 32. The two claims aren’t necessarily contradictory — 4K may describe the broader LTX VFX pipeline’s output ceiling rather than AlphaGen’s native resolution — but it’s worth testing before you plan a 4K matte-only workflow around it.
A hands-on test reported by Creative AI News found the model needs serious memory: a 49-frame clip at 3200×1900 used roughly 90GB of VRAM on an H100/B200-class card, well beyond the documented resolution ceiling, and took about 2.5 minutes to render. The same tester noted the model has no mask or prompt control — if it picks the wrong subject, you can’t correct it, only re-run and hope — and that content leaks into the matte past roughly 145 frames, so long clips need splitting, which can leave visible seams at the cut points.
Where this sits in a real pipeline
The honest way to read AlphaGen right now is as a beta checkpoint (Lightricks’ own filename tags it 0.9) rather than a production rotoscoping replacement. It’s most useful exactly where traditional keying already struggles — footage shot without a green screen, archival clips, anything with motion blur or fine detail like hair — and least useful anywhere you need per-shot control over exactly what counts as foreground. Licensing runs under the LTX-2 community license: free for most users, with a paid commercial license required above $10 million in annual revenue, per Lightricks’ terms.