LTX-2.5 Alpha Gen: video mattes without a green screen

LTX-2.5 Alpha Gen is an open adapter from Lightricks that takes an ordinary RGB video clip and returns a grayscale alpha matte, frame for frame, with no green screen, no mask and no text prompt. It runs on your own GPU. Sume does not host it: Sume's background removal route is an image route, so a hosted equivalent for video is not something Sume documents.

This page reads the Alpha Gen model card and the LTX-2 repository README, both read on 2026-10-02, and sets them next to what Sume's hosted routes do today.

What does LTX-2.5 Alpha Gen actually do?

The card describes an in-context LoRA on the LTX-2.5 22B base model. You feed it an RGB clip and it writes a matte video where white is foreground and black is background, aligned to the source frames. The card names hair, fur, smoke and semi-transparent material as the hard cases it is aimed at, and the intended use is compositing: isolate a person, animal, object or character, then place it over a new background.

• Input: an RGB video up to 1920x1088 and at most 145 frames.
• Output: a grayscale matte video, white for foreground and black for background.
• Settings the card requires: LoRA strength 1.0, an empty text prompt, and stage-1 inference at native resolution only.
• Base: LTX-2.5 22B, with the distilled version recommended.

What are the limits before you build on it?

The card is direct about three. A clip over 145 frames is out of range, and the card says RGB bleeding occurs past that length. You also cannot choose which object becomes the foreground; the model decides. And the hardware bar is high: the card points to H100 or B200 class GPUs for full-HD 145-frame clips.

Licensing is the other check. The card says the adapter is governed by the LTX-2 community license, with full terms in the GitHub repository. Read those terms for your revenue size before you ship it in a product; the LTX-2 license threshold post covers that one clause.

What does Sume offer for cutting out a subject?

Sume's cutout is Sume RMBG 1.0, an image route. Its published contract returns a mirrored PNG with alpha, and one billed removal covers one image. In the hosted MCP the tool is rmbg_create, listed with the other paid media tools in MCP tools and gates. There is no video matte route in those docs, and no EXR or alpha-channel video export.

You can run RMBG per frame: export frames, send each image, and recompose. That works for a short loop, but each frame is a separate billed removal, the matte is not temporally aware, and you own the frame extraction and the recombine. Alpha Gen is built for the video case; Sume's route is not.

When should you pick which?

Choose Alpha Gen when the job is video compositing with hair, smoke or glass, you have a GPU in the class the card names, and the licence fits. Choose Sume's RMBG when the asset is a still image, a product shot or a sticker, and you want a hosted call with a PNG back. If you only need a new background behind a person, prompt-based video editing on a hosted model is a different route; the background change post shows how that works and what to check.

Sume does not run LTX models, so there is no model id to send. The video catalog is whatever GET /v1/videos/models returns for your key, described in Video generation. Check it rather than assuming an open model is listed.