LTX-2: open audio and video generation
Added New
Our verdict
The most practical open video model to try if you want generation you control, from a company outside the big labs. It is heavy (22B-parameter weights), so expect a high-end GPU or the documented fp8 and offload flags, and start with the fast distilled pipeline. Weights ship under the LTX Community License, not an OSI licence: read its terms before commercial use.
- Who it's for
- Builders and studios with GPU access who want local text- or image-to-video
- Leverage
- Video clips with sound from your own hardware
- Effort
- High
- Date added
What it does
LTX-2 is a DiT-based audio-video foundation model from Lightricks with synchronized audio and video, keyframes, IC-LoRA control models and LoRA support. The repo provides pipelines run with uv, including a fast DistilledPipeline and a slower production-quality DFR pipeline, and LTX-2 is built into ComfyUI core. LTX-2.5 is the recommended model.
How builders use it
- Clone the repo, run uv sync, then start with the distilled pipeline.
- On smaller GPUs, add --quantization fp8-cast --offload cpu as the README suggests.
- Use it through ComfyUI if you already run ComfyUI workflows.
- Move to the DFR pipeline only once a prompt works, since it costs more time and VRAM.
What to expect
Short clips with matching audio for prototypes and social content, without per-second API fees.
Related discoveries
- ToolCreative & Media
ComfyUI
Open-source node-graph engine for image, video and audio generation that runs locally and exposes an API.
NewEffort: High#Image#Video
- ToolCreative & Media
Chatterbox TTS
Open-source text-to-speech models from Resemble AI with voice cloning, installable with pip.
NewEffort: Medium⧉ Prompt#Audio#GitHub repo
- ToolCreative & Media
Invoke (InvokeAI)
An open-source, locally run image studio with a canvas, inpainting and node workflows for open models.
NewEffort: Medium#Image#GitHub repo
- ToolCoding & Shipping Agents
Aider
Open-source AI pair programming in your terminal, with git commits built in.
Effort: Low#GitHub repo#Local / BYO model
Want more like this?
Practical discoveries, without the noise.