Lightricks releases LTX-2.5 with open weights and multi-shot video

Lightricks released LTX-2.5 on 11 August, with the weights on Hugging Face from the start.

One clip, several shots

It is a 22-billion-parameter diffusion transformer that generates video and audio together. The new trick is native multi-shot: one generation can cut between connected wide, medium and close shots and keep the people in them consistent. A Gemma 4 text encoder reads the prompt, and a new diffusion-based video decoder turns the result into frames.

Output

Clips default to 24 fps and reach 3840×2176 through spatial upscaling. The frame count has to be a multiple of eight plus one, and width and height divisible by 32. An optional duration predictor sizes the clip to the action described, and a prompt enhancer rewrites a rough prompt before generation.

Checkpoints

Lightricks publishes a trainable bf16 dev checkpoint, a distilled one that needs eight steps, and int8 and NVFP4 quantised versions for smaller or newer GPUs.

Licence

The LTX-2 Community License makes it free to use, commercially included, for organisations with less than $10 million in annual revenue; above that, Lightricks wants a paid agreement.

In ComfyUI

ComfyUI 0.32, out late the same day, supports LTX-2.5 locally and adds partner nodes for Lightricks' hosted version.