DINet: Deformation Inpainting Network for Realistic Face Visually Dubbing on High Resolution Video (AAAI2023)

Paper         demo video      Supplementary materials

🤔 How to achive this boost in inference latency?

To achieve this, several changes were implemented:

• Removed DeepSpeech and utilized wav2vec for instant feature extraction, leveraging the speed and power of torch.
• Trained a lightweight model to map the wav2vec features to DeepSpeech, maintaining the existing process.
• Enhanced frames extraction for improved speed.
• These adjustments contribute to a reduction of up to 60% in inference latency compared to the original implementation, all while maintaining quality.

Additionally, Docker has been introduced to facilitate faster, simpler, and more automated facial landmarks extraction.

Tested on:

• Ubuntu (18 and 20)
• Python version >= 3.9

📖 Prerequisites

To get started, follow these steps:

• Download the resources (asserts.zip) n Google drive. Unzip the file and place the directory in the current directory (./). This zip file includes the model for mapping wav2vec to deepspeech, beside all other models.

• For running inference or training, you have two options:

Option 1: Docker (most preferable)

Build the Docker image using the following command.

  docker build --target inference -t dinet . 

Option 2: Conda Environment

Set up a Conda environment by executing the following commands.

  #Create the virtual environment 
  conda create -n dinet python=3.9

  #Activate the environment
  conda activate dinet

  #Install teh requirements 
  pip install -r requirements.txt

🚀 Inference

Run inference with example videos:

docker run --rm --gpus 'device=0' -v $PWD:/app dinet python3 inference.py --mouth_region_size=256 --source_video_path=./asserts/examples/testxxx.mp4 --source_openface_landmark_path=./asserts/examples/testxxx.csv --driving_audio_path=./asserts/examples/driving_audio_xxx.wav --pretrained_clip_DINet_path=./asserts/clip_training_DINet_256mouth.pth  

The results are saved in ./asserts/inference_result.

Run inference with custom videos.

First, use the following Docker command to extract facial landmarks for your video. Replace input_video with the correct name of your video. The output file will be saved in the root directory.

  docker run --rm -v "$PWD:/mnt" -it algebr/openface -c "cp /mnt/input_video.mp4 /tmp/video.mp4 && build/bin/FeatureExtraction -f /tmp/video.mp4 -2Dfp -out_dir /tmp && cp /tmp/video.csv /mnt/input_video.csv" 

Run inference using the following command.

docker run --rm --gpus 'device=0' -v $PWD:/app -it dinet python3 inference.py --mouth_region_size=256 --source_video_path=<path_to_your_video>.mp4 --source_openface_landmark_path=<path_to_openface_output>.csv --driving_audio_path=<path_to_your_audio_file>.wav --pretrained_clip_DINet_path=./asserts/clip_training_DINet_256mouth.pth  

The results are saved in ./asserts/inference_result.

Or using Conda.

  #Activate the environment
  conda activate dinet

  # Run the command
  python3 inference.py --mouth_region_size=256 --source_video_path=<path_to_your_video>.mp4 --source_openface_landmark_path=<path_to_openface_output>.csv --driving_audio_path=<path_to_your_audio_file>.wav --pretrained_clip_DINet_path=./asserts/clip_training_DINet_256mouth.pth  

🧠 Training

First, you need to build the Docker image with the training dependencies as follows.

  docker build --target training -t dinet-training . 

This command will create the necessary Docker image for training with the specified dependencies.

Data Processing

We release the code of video processing on HDTF dataset. You can also use this code to process custom videos.

1. Downloading videos from HDTF dataset. Splitting videos according to xx_annotion_time.txt and do not crop&resize videos.

2. Resampling all split videos into 25fps and put videos into "./asserts/split_video_25fps". You can see the two example videos in "./asserts/split_video_25fps". We use software to resample videos. We provide the name list of training videos in our experiment. (pls see "./asserts/training_video_name.txt")

3. Using openface to detect smooth facial landmarks of all videos. Putting all ".csv" results into "./asserts/split_video_25fps_landmark_openface". You can see the two example csv files in "./asserts/split_video_25fps_landmark_openface".

4. Extracting frames from all videos and saving frames in "./asserts/split_video_25fps_frame". Run

   docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 data_processing.py --extract_video_frame --source_video_dir <PATH_TO_DATASET>

5. Extracting audios from all videos and saving audios in "./asserts/split_video_25fps_audio". Run

   docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 data_processing.py --extract_audio --source_video_dir <PATH_TO_DATASET>

6. Extracting deepspeech features from all audios and saving features in "./asserts/split_video_25fps_deepspeech". Run

   docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 data_processing.py --extract_deep_speech

7. Cropping faces from all videos and saving images in "./asserts/split_video_25fps_crop_face". Run

   docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 data_processing.py --crop_face

8. Generating training json file "./asserts/training_json.json". Run

   docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 data_processing.py --generate_training_json

Training models

We split the training process into frame training stage and clip training stage. In frame training stage, we use coarse-to-fine strategy, so you can train the model in arbitrary resolution.

Frame training stage.

In frame training stage, we only use perception loss and GAN loss.

1. Firstly, train the DINet in 104x80 (mouth region is 64x64) resolution. Run

   docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 train_DINet_frame.py --augment_num=32 --mouth_region_size=64 --batch_size=24 --result_path=./asserts/training_model_weight/frame_training_64

   You can stop the training when the loss converges (we stop in about 270 epoch).

2. Loading the pretrained model (face:104x80 & mouth:64x64) and train the DINet in higher resolution (face:208x160 & mouth:128x128). Run

   docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 train_DINet_frame.py --augment_num=100 --mouth_region_size=128 --batch_size=80 --coarse2fine --coarse_model_path=./asserts/training_model_weight/frame_training_64/xxxxxx.pth --result_path=./asserts/training_model_weight/frame_training_128

   You can stop the training when the loss converges (we stop in about 200 epoch).

3. Loading the pretrained model (face:208x160 & mouth:128x128) and train the DINet in higher resolution (face:416x320 & mouth:256x256). Run

   docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 train_DINet_frame.py --augment_num=20 --mouth_region_size=256 --batch_size=12 --coarse2fine --coarse_model_path=./asserts/training_model_weight/frame_training_128/xxxxxx.pth --result_path=./asserts/training_model_weight/frame_training_256

   You can stop the training when the loss converges (we stop in about 200 epoch). Keep in mind that you may need to adjust the batch size to start the training.

Clip training stage.

In clip training stage, we use perception loss, frame/clip GAN loss and sync loss. Loading the pretrained frame model (face:416x320 & mouth:256x256), pretrained syncnet model (mouth:256x256) and train the DINet in clip setting. Run

docker run --rm --gpus 'device=0' -it -v $PWD:/app dinet-training python3 train_DINet_clip.py --augment_num=3 --mouth_region_size=256 --batch_size=3 --pretrained_syncnet_path=./asserts/syncnet_256mouth.pth --pretrained_frame_DINet_path=./asserts/training_model_weight/frame_training_256/xxxxxx.pth --result_path=./asserts/training_model_weight/clip_training_256

You can stop the training when the loss converges and select the best model (our best model is at 160 epoch).

Acknowledge

The AdaAT is borrowed from AdaAT. The deepspeech feature is borrowed from AD-NeRF. The basic module is borrowed from first-order. Thanks for their released code.