LTX Alpha Gen pulls alpha from video without a green screen
Lightricks has released LTX Alpha Gen, a model that builds a per-frame alpha matte for video that has already been shot or generated. No green screen, manual masks or prompt is needed: the model isolates the main subject in the frame on its own, and it can go straight into comp. The main promise concerns the places that are hardest to key: hair, smoke, glass and motion blur.
There are two ways to work with the model. The open weights run on your own hardware at resolutions up to 1920×1088. The LTX cloud API processes video up to 4K and returns a finished cut-out with transparency.
The source frame stays untouched
Generative video tools can repaint the source, and when a shot has already been approved by the director or the client, changes to the RGB are unwelcome. In LTX Alpha Gen the source frame comes back exactly as it was approved, and the matte is computed alongside it, without re-rendering the shot. For a compositor that means the mask can be laid over the original plate.
The second problem the model addresses is semi-transparency. In classic keying, stray strands of hair, smoke and reflections in glass are usually either clipped off or drag the background along with them, and have to be finished by hand. The company claims Alpha Gen preserves those soft transitions:
“Hair, smoke, glass and motion blur keep soft edges, so the result can be sent straight to comp with no separate keying stage,” the LTX page states.
What the artist gets as output
The result depends on which route you use to access the model.
- Open weights. The model is posted on HuggingFace and runs in ComfyUI, via Python or inside a studio's own pipeline. The output is a single video: a greyscale alpha matte matching the source frame for frame. Combining it with the original footage is up to the artist. Maximum resolution is 1920×1088.
- API. The service returns two videos, frame-synchronous with the input. The first is the same grey matte, the second a finished cut-out with embedded transparency. There is no audio in the result. Resolution goes up to 4K. Processing can be built into an automated chain; the parameters are described in the API reference.
The specifics on formats come from the model's page on the fal.ai platform, where it is listed as LTX 2.5 Alpha Gen. The matte arrives as MP4, the cut-out as ProRes 4444 with straight alpha intended for compositing. The same platform has an IC-LoRA-based variant that returns only the grey matte. The model page on fal.ai is marked as allowing commercial use. It also lists the limits on the input file: no more than 200 MB and 3072 frames, long side up to 3840 pixels, short side up to 2160.
Where to get it and what it costs
The description, examples and download links are gathered on the LTX Alpha Gen page. In the LTX API, processing is billed per second of input video duration, and the rate depends on the short side of the frame: up to 1080 pixels — $0.10 per second, up to 1440 — $0.25, up to 2160 (that is, 4K) — $1.00. Video with a short side below 1080 pixels is billed at the 1080p rate. The legal terms of use for LTX models are set out in the LTX Model License Terms and Conditions (revision dated 12 November 2026). This is a contract between LTX and the customer. There are no separate terms for commercial work with the open Alpha Gen weights on the model page.
Who will find it useful and where the pitfalls are
The biggest winners are likely to be small teams and solo motion designers, for whom roto on a complicated shot means long manual work. The ability to run the model locally in ComfyUI also matters for studios that cannot send client footage to the cloud. On the facts known now, this approach has several limitations.
- The local version is capped at 1920×1088. 4K plates will apparently have to be either downscaled for the matte or processed through the cloud API, where the 4K rate is markedly more expensive.
- VRAM requirements, processing speed and the clip-length limit for the local version have not been published. How well the model holds up on long shots on an ordinary workstation is something you will have to check yourself.
- The cut-out from the API comes with straight alpha. In packages that work with premultiplied imagery it has to be interpreted correctly, otherwise dark or light fringes may appear along the edges.
- The API does not return audio. The track will have to be laid back in from the source.
- There are no independent quality comparisons yet with familiar tools such as Roto Brush, SAM masks or neural matting models. The claims about hair and smoke remain the company's statement.
The same LTX page also shows other narrowly specialised adapters built around the LTX models — for example Clean Plate IC-LoRA for removing people and cars from the frame and Golden Hour Relighting IC-LoRA for relighting daytime clips. Against that backdrop, Alpha Gen looks like part of a move toward individual post-production operations in addition to generating clips from scratch. Alpha, a clean plate and relighting make up a compositor's basic kit. If the quality holds up on real shots, models like this could become a standard link in ComfyUI pipelines alongside traditional compositing.
Which LTX adapter would you try first?
- Alpha Gen
- Clean Plate IC-LoRA