Lightricks released LTX-2.5 on August 11, the latest version of its open-weights video generation line, five months after LTX-2.3 made this site’s coverage for shipping synchronized video and audio in a single pass. The new release is a 22-billion-parameter asymmetric dual-stream diffusion transformer, and the headline additions are aimed squarely at people trying to cut something that resembles an actual scene rather than a single disconnected clip.
What’s new since 2.3
The biggest addition is native multishot generation — the model can now produce multiple connected shots that hold a consistent character, environment, lighting, and voice across cuts, instead of treating every generation as an isolated clip that has to be stitched together after the fact. Paired with that is auto duration, which predicts a clip’s length directly from the prompt rather than forcing a user to guess and set a frame count up front (frame counts still have to satisfy frames % 8 == 1, a quirk carried over from the underlying architecture).
On the output side, LTX-2.5 targets native 4K HDR, and adds what Lightricks calls Diffusion Fidelity Rendering (DFR) — an additional rendering pass meant to produce “cinema-grade EXR output” suitable for dropping into an actual production pipeline rather than just a shareable MP4. The video decoder was also replaced with a diffusion-based version in place of the prior VAE, which multiple write-ups describe as producing sharper faces, textures, and on-screen text than 2.3’s output — though on-screen text still isn’t reliably legible, per the model card’s own limitations section. A distilled variant is also fast: outlets report it generating a 10-second clip from a still image in roughly 6.8 seconds on Nvidia’s newest hardware.
Open weights, again
Lightricks is repeating the move that made LTX-2.3 notable in the first place: full open weights alongside API access, under the LTX-2.x Community License, which permits free commercial use for organizations under $10 million in annual revenue (larger companies need a paid agreement). The full repository — base transformer, distilled variant, quantized versions, text encoder, VAEs, upscalers, and LoRA add-ons — is live on Hugging Face at Lightricks/LTX-2.5, and ComfyUI shipped official workflow templates the same day, including through the ComfyUI-LTXVideo wrapper.
Why it matters beyond a version bump
Most frontier video labs still treat weights as the product. Lightricks treating them as a distribution channel is the more consequential story here, same as it was with 2.3 — it’s what lets independent artists, small studios, and tool builders fine-tune or self-host a model that would otherwise be locked behind a metered API. Multishot support specifically closes a gap that’s mattered to anyone using these models for anything longer than a single beat: continuity across cuts is exactly the kind of manual, frustrating stitching work that made earlier open video models feel more like a novelty generator than a filmmaking tool.