Lightricks /LTX-2

About LTX-2

LTX-2 is Lightricks' open model that makes video and matching audio together. You run it locally with Python commands on a CUDA GPU, not a hosted API. The weights are gated on Hugging Face and about 66 GiB. Developers and small studios who want to run it themselves are the target.

Best for: Small studios and developers with a CUDA GPU who want to run the LTX-2 audio-video model themselves.

• Makes video and audio in one pass from a text or image prompt
• Ships CLI pipelines: distilled, DFR, IC-LoRA, Dub-It, TI2Vid, keyframe, A2Vid
• Runs on less VRAM with --quantization fp8-cast --offload cpu
• Defaults to 1024x1536 at 24 fps; 4K is --width 3840 --height 2176
• Uses a bundled Gemma 4 12B text encoder, so no separate download

1. 1
   Clone the repo: the command below then cd LTX-2

   git clone https://github.com/Lightricks/LTX-2.git

   Copy
2. 2
   Install dependencies with uv sync --extra natten (the natten extra is Linux + CUDA only and is skipped elsewhere)
3. 3
   Log in with hf auth login, then download the LTX-2.5 weights with the command below (about 66 GiB)

   hf download Lightricks/LTX-2.5 ... --local-dir models/ltx-2.5

   Copy
4. 4
   Run the fast distilled pipeline: the command below

   uv run python -m ltx_pipelines.distilled --transformer-path ... --output-path output.mp4 --prompt "..."

   Copy
5. 5
   For production quality, download the detailing IC-LoRA and run the command below with --detailing-lora

   uv run python -m ltx_pipelines.dfr_pipeline

   Copy
6. 6
   If VRAM is tight, add the command below

   --quantization fp8-cast --offload {cpu, disk}

   Copy