LTX-2.5
Lightricks' 22B open-weights model generating synchronized video and audio with multi-shot consistency.
About
Video and audio come out of LTX-2.5 together: Lightricks' 22B parameter generator produces footage with synchronized sound from text, image, or video conditioning, and its multishot mode holds character identity and visual style stable across consecutive scenes. The 2026 release ships as a full trainable transformer plus a faster distilled variant, with int8 and NVFP4 quantizations, a choice of two VAEs trading quality against speed, and FP8 casting with CPU offload for constrained GPUs; generation runs at 24 fps in clips of up to 121 frames or more per pass. Official workflow templates land in ComfyUI on day one, which has made the LTX line a staple of the open video community, and the stack expects Python 3.12, CUDA 12.7, and PyTorch 2.7. Weights are distributed under the LTX-2 Community License Agreement: commercial use is free for organizations under 10 million dollars in annual revenue, with paid licensing above that line, so hobbyists and small studios run it without cost while large companies negotiate terms.
Should you use LTX-2.5?
Pick it when
Pick LTX-2.5 when you want picture and synchronized sound in one pass and characters that stay consistent across several shots, with day-one ComfyUI templates, and your organization is under 10 million dollars in annual revenue.
Look elsewhere when
Skip it if your company is above the 10 million dollar revenue line and will not pay for a license, or you need fully permissive weights: Ovi does audio-video under Apache-2.0. It also expects a recent Python 3.12 and CUDA 12.7 stack.
Alternatives to LTX-2.5
- Ovi
Apache-2.0 video plus speech and effects in one pass at 24 GB with FP8, but clips are 5 or 10 seconds with no multishot identity mode.
- MiniMax H3
Larger 33B omni-modal model with native stereo audio, speech in 11 languages, and audio or video conditioning, but multi-GPU for local use.
- LTX-Video
Older LTX checkpoints, including a 2B distilled model and FP8 variants, lighter for silent clips on modest GPUs and macOS.
- SkyReels-V3
Reference-to-video keeps 1 to 4 specific people or products consistent and includes an audio-driven avatar model, under the Skywork Community License.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Video Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- LTX-2 Community License
- Added
- Aug 24, 2026