LaVie: High-Quality Video Generation with Cascaded Latent Diffusion Models

This repository is the official PyTorch implementation of LaVie.

LaVie is a Text-to-Video (T2V) generation framework, and main part of video generation system Vchitect. You can also check our fine-tuned Image-to-Video (I2V) model SEINE.

Installation

conda env create -f environment.yml 
conda activate lavie

Download Pre-Trained models

Download pre-trained LaVie models, Stable Diffusion 1.4, stable-diffusion-x4-upscaler to ./pretrained_models. You should be able to see the following:

├── pretrained_models
│   ├── lavie_base.pt
│   ├── lavie_interpolation.pt
│   ├── lavie_vsr.pt
│   ├── stable-diffusion-v1-4
│   │   ├── ...
└── └── stable-diffusion-x4-upscaler
        ├── ...

Gallery:


two teddy bears playing poker under water, highly detailed, oil painting style	a teddy bear skateboarding under water, highly detailed	a cat reading a book on the table, Van Gogh style

a cute raccoon playing guitar in the park at sunrise, oil painting style	a teddy bear walking in the park at sunrise, oil painting style	a teddy bear reading a book near a small river, oil painting style

Elon Musk in a space suit standing besides a rocket, high quality	a teddy bear in a suit having dinner in a well-decorated house	Iron Man flying in the sky, 4k, high quality

Feel free to try different prompts, and share with us which one you like the most!

Inference

The inference contains Base T2V, Video Interpolation and Video Super-Resolution three steps. We provide several options to generate videos:

	Step1	Step2	Step3	Resolution	Length
option1	✔			320x512	16
option2	✔	✔		320x512	61
option3	✔		✔	1280x2048	16
option4	✔	✔	✔	1280x2048	61

Feel free to try different options :)

Step1. Base T2V

Run following command to generate videos from base T2V model.

cd base
python pipelines/sample.py --config configs/sample.yaml

In configs/sample.yaml, arguments for inference:

ckpt_path: Path to the downloaded LaVie base model, default is ../pretrained_models/lavie_base.pt
pretrained_models: Path to the downloaded SD1.4, default is ../pretrained_models
output_folder: Path to save generated results, default is ../res/base
seed: Seed to be used, None for random generation
sample_method: Scheduler to use, default is ddpm, options are ddpm, ddim and eulerdiscrete
guidance_scale: CFG scale to use, default is 7.5
num_sampling_steps: Denoising steps, default is 50
text_prompt: Prompt for generation

Following results were generated with the arguments:

seed: 400, sample_method: ddpm, guidance_scale: 7.0, num_sampling_steps: 50

(you might obtain different results on different device)


a Corgi walking in the park at sunrise, oil painting style	a panda taking a selfie, 2k, high quality	a polar bear playing drum kit in NYC Times Square, 4k, high resolution

a shark swimming in clear Carribean ocean, 2k, high quality	a teddy bear walking on the street, 2k, high quality	jungle, river, at sunset, ultra quality

Step2 (optional). Video Interpolation

Run following command to conduct video interpolation.

cd interpolation
python sample.py --config configs/sample.yaml

The default input video path is ./res/base, results will be saved under ./res/interpolation. In configs/sample.yaml, you could modify default input_folder with YOUR_INPUT_FOLDER in configs/sample.yaml. Input videos should be named as prompt1.mp4, prompt2.mp4, ... and put under YOUR_INPUT_FOLDER. Launching the code will process all the input videos in input_folder.


a teddy bear walking on the street, 2k, high_quality	a Corgi walking in the park at sunrise, oil painting style	a panda taking a selfie, 2k, high quality

Step3 (optional). Video Super-Resolution

Run following command to conduct video super-resolution.

cd vsr
python sample.py --config configs/sample.yaml

The default input video path is ./res/base and results will be saved under ./res/vsr. You could modify default input_path with YOUR_INPUT_FOLDER in configs/sample.yaml. Similar to Step2, input videos should be named as prompt1.mp4, prompt2.mp4, ... and put under YOUR_INPUT_FOLDER. Launching the code will process all the input videos in input_folder.

BibTex

@article{wang2023lavie,
  title={LAVIE: High-Quality Video Generation with Cascaded Latent Diffusion Models},
  author={Wang, Yaohui and Chen, Xinyuan and Ma, Xin and Zhou, Shangchen and Huang, Ziqi and Wang, Yi and Yang, Ceyuan and He, Yinan and Yu, Jiashuo and Yang, Peiqing and others},
  journal={arXiv preprint arXiv:2309.15103},
  year={2023}
}

@article{chen2023seine,
title={SEINE: Short-to-Long Video Diffusion Model for Generative Transition and Prediction},
author={Chen, Xinyuan and Wang, Yaohui and Zhang, Lingjun and Zhuang, Shaobin and Ma, Xin and Yu, Jiashuo and Wang, Yali and Lin, Dahua and Qiao, Yu and Liu, Ziwei},
journal={arXiv preprint arXiv:2310.20700},
year={2023}
}

Disclaimer

We disclaim responsibility for user-generated content. The model was not trained to realistically represent people or events, so using it to generate such content is beyond the model's capabilities. It is prohibited for pornographic, violent and bloody content generation, and to generate content that is demeaning or harmful to people or their environment, culture, religion, etc. Users are solely liable for their actions. The project contributors are not legally affiliated with, nor accountable for users' behaviors. Use the generative model responsibly, adhering to ethical and legal standards.

Contact Us

Yaohui Wang: wangyaohui@pjlab.org.cn
Xinyuan Chen: chenxinyuan@pjlab.org.cn
Xin Ma: xin.ma1@monash.edu

Acknowledgements

The code is built upon diffusers and Stable Diffusion, we thank all the contributors for open-sourcing.

License

The code is licensed under Apache-2.0, model weights are fully open for academic research and also allow free commercial usage. To apply for a commercial license, please contact vchitect@pjlab.org.cn.

Vchitect / LaVie