How much memory does it need?

If GPU memory runs out, append `--quantization fp8-cast` to the generation command and add either `--offload cpu` or `--offload disk`. For example, put a `\` after the final prompt line above and add the option on the next line. CPU or disk offload changes the execution path and runtime. The reported 6.8 seconds for a 10-second 720p clip on 2×GB200 GPUs applies to that server configuration. [LTX-2.5 announcement](https://ltx.io/model/ltx-2-5)

Image and video generation

Install LTX-2.5 locally and generate your first video

Start with the 22B distilled checkpoint and the 121-frame command.

LTX-2.5 is an open-weight model that generates video with synchronized audio from a prompt¹. The official LTX-2 repository provides commands to install its Python pipeline², download roughly 66 GiB of checkpoint³ files, and save output.mp4 with the distilled pipeline.

Prepare Python and model storage before installation

LTX-2.5 is an open-weight model that generates video from a text prompt along with audio matched to the scene. This guide uses the official LTX-2 Python pipeline, downloads the distilled checkpoint, and saves one 121-frame MP4. Official repository

The official setup uses Python and uv, adding the fast natten extension only on Linux. Install it using the official uv instructions and confirm uv --version prints a version. On Linux, include --extra natten in the repository setup command below; on Windows and macOS, omit it and use the fallback decoder. Model files total about 66 GiB. Accept the LTX-2.5 terms on Hugging Face and log in with a Read token⁴ before downloading.

Clone the repository and install dependencies

uv --version
git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2
# Linux only; omit --extra natten on Windows and macOS
uv sync --extra natten

A storyboard showing a cyclist on a coastal road across several frames — Choose frame count and aspect ratio to produce one clip.

Download the five required checkpoint files

Accept the terms for Lightricks/LTX-2.5 on Hugging Face, then register a Read token with hf auth login. The download keeps the repository’s folder structure beneath --local-dir, matching the generation command. Reserve at least 66 GiB before downloading, then confirm that each file exists. Official model download and paths

Download the required files from Hugging Face

hf auth login
hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  vae/ltx-2.5-audio-vae-bf16.safetensors \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir models/ltx-2.5

Generate your first MP4 with the distilled pipeline

Once the five files are saved, run the official ltx_pipelines.distilled command. --num-frames 121 sets the frame⁵ count for this request, and --seed 42 helps make runs easier to compare. Describe the scene, motion, camera, and sound in the prompt. The command does not set a playback frame rate, so this guide does not assign a duration.

Generate the MP4

uv run python -m ltx_pipelines.distilled \
  --transformer-path models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
  --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --num-frames 121 \
  --seed 42 \
  --output-path output.mp4 \
  --prompt "A small wooden sailboat glides across a calm lake at sunrise, camera follows from the shore, gentle water sounds and distant birds."

Three film reels of different sizes and stacks of landscape contact prints on a wooden desk. — Plan for the storage and memory used by frames and model files.

Check the output and the first failure point

When the command finishes, open output.mp4 in the current run directory and confirm that the video plays and has an audio track. If the first run stops, inspect the first traceback in the terminal. For 401 or 403, check Hugging Face acceptance and token permissions; for a missing-file error, compare the path and name under models/ltx-2.5/.

If GPU⁶ memory runs out, append --quantization fp8-cast to the generation command and add either --offload cpu or --offload disk. For example, put a \ after the final prompt line above and add the option on the next line. CPU⁷ or disk offload⁸ changes the execution path and runtime⁹. The reported 6.8 seconds for a 10-second 720p clip on 2×GB200 GPUs applies to that server configuration. LTX-2.5 announcement

Official LTX-2.5 setup and generation commands

An editing work surface showing a strip of video scenes with audio waveforms — LTX-2.5 also generates audio synchronized with the video scene.

Terminology notes

Prompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the text
Pipeline — A sequence of processing stages from input to output. Different models or tools may be used at each stage.
Back to the text
Checkpoint — A file containing saved model weights and related state. Versions or tasks in one model family may use different checkpoints.
Back to the text
Token — A unit into which a model divides input or output for processing.
Back to the text
Frame — A single image that makes up part of a video.
Back to the text
GPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the text
CPU — The central processor that runs the operating system and general-purpose instructions; in AI workloads it also handles tasks such as preprocessing and data movement.
Back to the text
Offloading — Moving some model data from GPU memory to system RAM or storage when capacity is limited. This adds data transfer.
Back to the text
Runtime — The software environment that provides facilities needed while a program runs. In local AI, it can also refer to a model execution engine.
Back to the text

Frequently asked questions

What should you check first in this article?: LTX-2.5 is an open-weight model that generates video with synchronized audio from a prompt. The official LTX-2 repository provides commands to install its Python pipeline, download roughly 66 GiB of checkpoint files, and save `output.mp4` with the distilled pipeline.
How much memory does it need?: If GPU memory runs out, append `--quantization fp8-cast` to the generation command and add either `--offload cpu` or `--offload disk`. For example, put a `\` after the final prompt line above and add the option on the next line. CPU or disk offload changes the execution path and runtime. The reported 6.8 seconds for a 10-second 720p clip on 2×GB200 GPUs applies to that server configuration. [LTX-2.5 announcement](https://ltx.io/model/ltx-2-5)