LTX-2.5

LTX-2.5 is the August 2026 release from LTX, the video-model company spun out of Lightricks. It is a 22-billion-parameter diffusion-transformer model that generates synchronized audio and video in one pass, distributed as open weights through a gated Hugging Face repository with ComfyUI integration on day one. The headline feature is native multishot generation: several connected shots with consistent characters, voice, and lighting produced in a single inference pass instead of stitched-together clips. The 2.5 upgrade keeps the 22B class but reworks the stack around it: a diffusion-based video decoder for sharper faces and text, a custom Gemma 4 12B text encoder with projection, an optional duration head that predicts clip length from the prompt, and a retrained distilled Fast checkpoint. LTX reports 10 seconds of image-to-video in 6.8 seconds on NVIDIA superchips.

Licensing is custom and important to read: the LTX-2.x Community License permits use free of charge for organizations under $10M in annual recurring revenue; larger companies negotiate a paid license, and the license does not apply in the EU, UK, or South Korea per third-party analysis. LTX says the model family has passed 33 million downloads, making it one of the most-used open video-model lines. Beyond generation, LTX-2.5 edits real footage and exports to EXR for film and VFX pipelines, which positions it as much as a post-production tool as a generator.