🌅 DAWN: Dynamic Frame Avatar with Non-autoregressive Diffusion Framework for Talking Head Video Generation

😊 Please give us a star ⭐ to support us for continous update 😊

News

2024.10.14 🔥 We release the Demo page.
2024.10.18 🔥 We release the paper DAWN.
2024.10.21 🔥 We update the Chinese introduction .
2024.11.7 🔥🔥 We realse the pretrained model on hugging face.
2024.11.9 🔥🔥🔥 We realse the inference code. We sincerely invite you to experience our model. 😊

TODO list:

release the inference code
release the pretrained model of 128*128
release the pretrained model of 256*256
release the test code for HDTF dataset
in progress ...

Equipment Requirements

With our VRAM-oriented optimized code, the maximum length of video that can be generated is linearly related to the size of the GPU VRAM. Larger VRAM produce longer videos.

To generate 128*128 video, we recommend using a GPU with 12GB or more VRAM. This can at least generate video of approximately 400 frames.
To generate 256*256 video, we recommend using a GPU with 24GB or more VRAM. This can at least generate video of approximately 200 frames.

PS: Although optimized code can improve VRAM utilization, it currently sacrifices inference speed due to incomplete optimization of local attention. We are actively working on this issue, and if you have a better solution, we welcome your PR. If you wish to achieve faster inference speeds, you can use unoptimized code, but this will increase VRAM usage (O(n²) spatial complexity).

Methodology

The overall structure of DAWN:

Environment

We highly recommend to try DAWN on linux platform. Runing on windows may produce some rubbish files need to be deleted manually and requires additional effort for the deployment of the 3DDFA repository (our extract_init_states folder) comment.

set up the conda environment

conda create -n DAWN python=3.8
conda activate DAWN
pip install -r requirements.txt
conda create -n 3DDFA python=3.8
conda activate 3DDFA
pip install -r requirements_3ddfa.txt

Follow the readme and 3DDFA_V2 to set up the 3DDFA environment.

Inference

Since our model is trained only on the HDTF dataset and has few parameters, in order to ensure the best driving effect, please provide examples of :

standard human photos as much as possible, try not to wear hats or large headgear
ensure a clear boundary between the background and the subject
have the face occupying the main position in the image.

The preparation for inference:

Download the pretrain checkpoints from hugging face. Create the ./pretrain_models directory and put the checkpoint files into it.
Changing the path in run_ood_test\run_DM_v0_df_test_128_both_pose_blink.sh or run_ood_test\run_DM_v0_df_test_256_1.sh. Infill the image_path, audio_path and cache_path. The run_ood_test\run_DM_v0_df_test_128_both_pose_blink.sh is used to perform inference on 128 * 128 images and run_ood_test\run_DM_v0_df_test_256_1.sh is used to perform inference on 256 * 256 images.
Using bash xxxx.sh to run the script.

About PBNet

We provide two PBNet checkpoints: 1. generating both blink and pose together 2. generating blink and pose separately (script end with "separate_pose_blink"). According to the quantitative results, these two methods have similar performance.

This code is tested on internal server of company and my Windows 11 PC. There might be some minor problems due to the difference of the equipment. Please feel free to leave issues or PR if you encounter some problems, we are glad to help!

For testing on datasets: If you wish to test the performance of DAWN on datasets, we recommend to warp our code and process data in batches for each step (including extracting the initial states, audio embedding, inference of PBNet, inference of A2V-FDM). Reloading the model repeatedly will make your testing efficiency very low. We also plan to release the inference code for HDTF dataset in the future.

Citing DAWN

If you wish to refer to the baseline results published here, please use the following BibTeX entries:

@misc{dawn2024,
      title={DAWN: Dynamic Frame Avatar with Non-autoregressive Diffusion Framework for Talking Head Video Generation}, 
      author={Hanbo Cheng and Limin Lin and Chenyu Liu and Pengcheng Xia and Pengfei Hu and Jiefeng Ma and Jun Du and Jia Pan},
      year={2024},
      eprint={2410.13726},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2410.13726}, 
}

Acknowledgement

Limin Lin and Hanbo Cheng contributed equally to the project.

Thank you to the authors of Diffused Heads for assisting us in reproducing their work! We also extend our gratitude to the authors of MRAA, LFDM, 3DDFA_V2 and ACTOR for their contributions to the open-source community. Lastly, we thank our mentors and co-authors for their continuous support in our research work!

Hanbo-Cheng/DAWN-pytorch