Blog ·

LTX-2.5 online: try Lightricks' open video model with sound, no 80 GB GPU needed

LTX-2.5 generates video with synchronised audio from text or an image. What's new, its license (free under $10M revenue), how to try it in your browser and how to run it locally.

LTX-2.5 is the latest open video model from Lightricks: a 22-billion-parameter model that generates video together with its sound (voices, effects, ambience) from a text prompt or a starting image. The weights are free to download, but running them needs a GPU with 80 GB or more, which almost nobody has at home. On LiveWeights you can try it in your browser.

LTX-2.5Lightricks · LTX-2.x Community License (free under $10M annual revenue)Lightricks' open 22B video model: clips with synchronised sound from a text prompt or a starting image, in landscape, portrait or square.Try it online →

What's new in LTX-2.5

  • Video and audio in one pass: dialogue, sound effects and background ambience are generated with the picture and stay in sync.
  • Multishot scenes: connected shots that keep the same characters, lighting and style across cuts.
  • A new diffusion video decoder: sharper faces, textures and on-screen text than the previous VAE decoder.
  • A much better distilled model: 8 steps plus a 2× upscaler, close to the full model's quality at a fraction of the time. This is the version we host.

License: who can use it commercially?

LTX-2.5 is released under the LTX-2.x Community License: commercial and production use are free for companies and creators with less than $10 million in annual revenue. Above that, Lightricks asks for a paid license. Videos you generate on LiveWeights are yours to use under those terms.

How to try it online

  1. Open the LTX-2.5 page and sign in, then pick a small credit pack (no subscription).
  2. Describe the scene: what we see, how the camera moves, and what we hear. Add a starting image if you want to animate a photo or an illustration.
  3. Pick landscape (16:9), portrait (9:16, for Reels and Shorts) or square, and 5 or 10 seconds.
  4. Press Generate: a 5-second clip takes about a minute or two on our 96 GB GPU, longer if the model has to load first.

Tips for better results

  • Write like a director: subject, action, camera ("slow dolly in", "handheld", "aerial"), light, and sound.
  • Put spoken lines in quotes: the model generates the voice and the lip movement.
  • For image-to-video, describe the motion, not the picture: the image already shows what things look like.

Run it locally

The weights are on Hugging Face (Lightricks/LTX-2.5, gated: accept the license first). The official Python package is ltx-pipelines from the LTX-2 GitHub repository; ComfyUI ships official workflows, and a Diffusers version is in the works. The bf16 distilled model needs about 80 GB of GPU memory; with fp8 and CPU offload it fits on 32–48 GB cards, more slowly. For another open video model, try Hy-Nomas, Tencent's interactive world model.

LiveWeights is not affiliated with the model authors or with the AI Search channel. Facts come from the official pages linked on each model page.