LTX-Video-xDiT

This project is based on the LTX-Video algorithm of the diffusers and optimized and accelerated for multi GPUs inference using the xDiT framework.

Installation

To set up the environment and install the required dependencies, follow these steps:

1. Create and activate the Conda environment:

conda create -n ltxvideo-xfuser python=3.10
conda activate ltxvideo-xfuser

2. Install PyTorch and related libraries:

pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu124

3. Install the xDiT package:

git clone https://github.com/zishen-ucap/LTX-Video-xDiT.git
cd LTX-Video-xDiT
pip install -e .

You can install flash-att==2.6.3 by this link:

flash-attn==2.6.3

pip install flash_attn-2.6.3+cu123torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

4. Download the model parameters:

You can download the model parameters from this link:

Lightricks/LTX-Video

Alternatively, you can use the following command to download them:

pip install -U huggingface_hub
huggingface-cli download --resume-download Lightricks/LTX-Video --local-dir Lightricks/LTX-Video --local-dir-use-symlinks False

Usage

Once the setup is complete, you can run the example with the following command:

CUDA_VISIBLE_DEVICES=0,1,2,3, torchrun --nproc_per_node=4 examples/ltxvideo_example.py --model './Lightricks/LTX-Video' --ulysses_degree 1 --ring_degree 2 --use_cfg_parallel --height 512 --width 768 --num_frames 161 --prompt "Your prompt" --num_inference_steps 50 --seed 42