LTX-2.5

LTX-2.5 is Lightricks' open-weight video model that makes synced video and audio from text, images or video, for creators who run models locally.

What it is, and what is new in 2.5

LTX-2.5 is a single DiT-based model with open weights, built for local execution and fine-tuning. The file names point to a 22B-parameter transformer. The docs describe it as an "open world model" whose established use is video with sound. Its use in robotics and physical AI is described as still developing.

• Native multishot: one generation can produce several connected shots that keep character, scene, lighting, style and voice across cuts. Earlier versions made a single continuous shot.
• Auto duration: an optional duration head predicts the clip length from the prompt instead of a fixed setting.
• A new diffusion video decoder, which Lightricks says improves faces, textures, typography and fast motion, and a custom Gemma 4 12B text encoder for complex prompts.
• An optional prompt enhancer that expands a short prompt, and which you can turn off to keep your exact wording.
• Native EXR input and output, so footage does not have to go through a lossy 8-bit round trip.
• An improved distilled model that Lightricks says keeps more of the full model's quality at much lower compute. These are the company's own descriptions, not Seailife tests.

Ways to run it

• Text-to-video and image-to-video, with the opening frame set by one or more images.
• Audio-to-video in ComfyUI: your clip is encoded, held fixed, and the video is generated to follow its pacing. The output MP4 carries your original audio unchanged.
• Text-to-audio only, with no video branch.
• IC-LoRA adapters for control, edit and restore tasks. The product page lists Refine, Restore, SDR to HDR, HDR native from 16-bit EXR, Layout to Render, and a beta AlphaGen for soft alpha mattes.
• Standard LoRAs. Lightricks says most LoRAs and IC-LoRAs trained on LTX-2.3 run on 2.5 unchanged, with a few exceptions to validate. No general-purpose standard-LoRA workflow for 2.5 is published yet.
• Foley for muted video and Dub-It for replacing dialogue are still LTX-2.3 workflows, and the docs warn not to swap in 2.5 assets without compatible-workflow evidence.

Getting it running

• The docs recommend ComfyUI first, with ready-made LTX-2.5 templates. For code, the ltx-pipelines Python package works from the Lightricks/LTX-2 repo, and a Diffusers pack is at Lightricks/LTX-2.5-Diffusers. Diffusers support is not in a release yet, so you install it from the main branch.
• Python 3.12 or later, CUDA 12.7 or later and PyTorch around 2.7 are recommended for the Python route.
• The weights are a split pack with one safetensors file per component. They are gated: log in with hf auth login and click Agree and Access on the Hugging Face page first.
• The distilled transformer uses a fixed 8-step schedule with CFG set to 1. Frame count must satisfy frames % 8 == 1 and width and height must be divisible by 32. The distilled flow runs a first stage and then a 2x spatial upscale and refine pass.
• For less VRAM, the Hugging Face page suggests --quantization fp8-cast with --offload cpu. The int8 files are ComfyUI only, and the NVFP4 file needs a Blackwell GPU through ltx-pipelines or ComfyUI.
• LTX's own comparison table claims a 16GB minimum VRAM and faster than real time. That is the company's claim. Seailife did not time the model on any card.

License: free under 10 million dollars in revenue

• The weights fall under the LTX-2.x Community License. Commercial and production use is free for entities under 10 million US dollars in annual revenue, counted across the whole entity including subsidiaries and affiliates under common control.
• Transfer of fine-tunes may require a paid license, as the license sets out. The Hugging Face page says the full binding terms are in its LICENSE file.
• LTX describes the licensing as permissive with no mandatory branding. Its comparison table contrasts that with MiniMax H3, which it marks as open with conditions, not available in the US, EU, UK and Korea, and requiring the "MiniMax H3" name to be shown. You can read the other side on the MiniMax H3 page.
• The Lightricks/LTX-2 code repository on GitHub is Apache 2.0 licensed, with about 10,100 stars, and was last pushed on 2026-11-05. The weights have a separate license from the code.

Hosted options, and limits to know

• You can try it in the LTX API playground without installing anything, and there is an API with a key. Seailife did not check API prices.
• Lightricks' own limitations list says the model is not meant to give factual information, can amplify societal biases, may produce inappropriate or offensive content, may not match the prompt exactly, and follows prompts differently depending on style.
• The Hugging Face page shows about 10,900 likes. Seailife has not run the model.