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

Paper demo video Supplementary materials

Inference

Download resources (asserts.zip) in Google drive. unzip and put dir in ./.

Inference with example videos. Run

python 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

Inference with custom videos.
Note: The released pretrained model is trained on HDTF dataset with 363 training videos (video names are in ./asserts/training_video_name.txt), so the generalization is limited. It would be better to test custom videos with normal lighting, frontal view etc.(see the limitation section in the paper). We also release the training code, so if a larger high resolution audio-visual dataset is proposed in the further, you can use the training code to train a model with greater generalization. Besides, we release coarse-to-fine training strategy, so you can use the training code to train a model in arbitrary resolution (larger than 416x320 if gpu memory and training dataset are available).

Using openface to detect smooth facial landmarks of your custom video. We run the OpenFaceOffline.exe on windows 10 system with this setting:

Record	Recording settings	OpenFace setting	View	Face Detector	Landmark Detector
2D landmark & tracked videos	Mask aligned image	Use dynamic AU models	Show video	Openface (MTCNN)	CE-CLM

The detected facial landmarks are saved in "xxxx.csv". Run

python inference.py --mouth_region_size=256 --source_video_path= custom video path --source_openface_landmark_path=  detected landmark path --driving_audio_path= driving audio path --pretrained_clip_DINet_path=./asserts/clip_training_DINet_256mouth.pth

to realize face visually dubbing on your custom videos.

Training

Data Processing

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

Downloading videos from HDTF dataset. Splitting videos according to xx_annotion_time.txt and do not crop&resize videos.
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")
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".
Extracting frames from all videos and saving frames in "./asserts/split_video_25fps_frame". Run

python data_processing.py --extract_video_frame

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

python data_processing.py --extract_audio

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

python data_processing.py --extract_deep_speech

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

python data_processing.py --crop_face

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

python 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.

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

python 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).

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

python 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).

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

python 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).

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

python 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/xxxxx.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.

jack20002/DINet-