Practical overview
When it fits
Use LTX-2 when you want a documented Python/CLI workflow for audio-video generation or LoRA training, and can manage model downloads and local compute resources.
Who it is for
- Developers who want to run audio-video generation from a Python command-line workflow.
- Practitioners who want to train LoRAs for the LTX-2 model.
- Users comparing a faster distilled generation path with a higher-quality DFR rendering path.
Documented uses
- Generate an audio-video output from a text prompt and save it as an MP4.
- Prepare model weights and dependencies for local inference.
- Add a detailing IC-LoRA and use the documented DFR route for text- or image-to-video rendering.
- Train LoRAs to adapt the generative model.
Requirements & limits
The repository and model files are separate: the README lists roughly 66 GiB of model downloads and says gated model access may require accepting terms and logging in with a Hugging Face read token.
- The quick-start model files total roughly 66 GiB according to the README.
- DFR is described as slower and requiring more VRAM than the distilled pipeline.
- The optional
nattenbackend is limited to Linux with CUDA; the README says Windows and macOS use a Triton or eager fallback. - The documented setup uses uv, the Hugging Face CLI, Python pipeline modules, and locally referenced model files.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up and generate with the distilled pipeline
Install LTX-2, download its listed model files, and run the documented prompt-to-MP4 inference command.
Before you start
- A cloned LTX-2 repository and uv for dependency installation.
- Hugging Face CLI access to the LTX-2.5 model files; accepting model terms and a read token may be required.
- Roughly 66 GiB for the listed model files.
- Python dependencies installed with the documented uv command.
Steps
Clone the repository and enter its directory.
Install dependencies with the command below.
uv sync --extra nattenLog in with the Hugging Face CLI and download the listed transformer, text encoder, VAEs, and spatial upscaler into
models/ltx-2.5.Run the command below with the documented model paths, frame count, seed, output path, and text prompt.
uv run python -m ltx_pipelines.distilled
Expected outcome
The example pipeline is configured to write a generated audio-video result to an MP4 output path.
Guide limitations
- The
nattenextra is Linux and CUDA only; the README says Windows and macOS skip it and use Triton or eager decoding. - The README describes roughly 66 GiB for the listed files and says GPU-memory constraints may call for FP8 casting or CPU/disk offload.
- This recipe reflects documented commands and is not a report of an execution test.
Prepare the DFR rendering model files
Add the detailing IC-LoRA used by the documented DFR pipeline to the files from the distilled setup.
Before you start
- The text encoder, VAEs, spatial upscaler, and distilled transformer from the initial model download.
- Hugging Face CLI access to the detailing IC-LoRA repository; model access may require accepting terms and a read token.
Steps
Download the detailing IC-LoRA file from
Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscalerintomodels/ltx-2.5/loras.Use that file with the DFR pipeline as described in the repository documentation.
Expected outcome
The added detailing IC-LoRA and reused distilled model files prepare the documented DFR path for text- or image-to-video rendering.
Guide limitations
- DFR is described as slower and requiring more VRAM than the distilled starting pipeline.
- The repository says to use the distilled transformer, not the full development transformer.
- This recipe covers model-file preparation; it does not claim an execution test.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- Lightricks/LTX-2 · Official homepage title · retained in its source language