TalkLip net

This repo is the official implementation of 'Seeing What You Said: Talking Face Generation Guided by a Lip Reading Expert', CVPR 2023.

Arxiv | Paper

🔥 News

We upload a Talking_face_demo.pptx to this repository which contains some demo videos.
Fix the GPU out-of-memory error in train.py. Running train.py with a batch_size of 8 requires approximately 24GB of memory. However, in some rare cases, it might need more than 24GB and trigger an error. We have resolved this issue using a try-and-catch mechanism. -- 19/July/2023
We upload a checkpoint of the discriminator as requested in the issue.
We upload an eval_lrs.sh in the evaluation folder, which allows you to evaluate all metrics on LRS2 with a single command.

Prerequisite

pip install torch==1.12.0+cu113 torchvision==0.13.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html.
Install AV-Hubert by following his installation.
Install supplementary packages via pip install -r requirement.txt
Install ffmpeg. We adopt version=4.3.2. Please double check wavforms extracted from mp4 files. Extracted wavforms should not contain prefix of 0. If you use anaconda, you can refer to conda install -c conda-forge ffmpeg==4.2.3
Download the pre-trained checkpoint of face detector pre-trained model and put it to face_detection/detection/sfd/s3fd.pth. Alternative link.

Dataset and pre-processing

Download LRS2 for training and evaluation. Note that we do not use the pretrain set.
Download LRW for evaluation.
To extract wavforms from mp4 files:

python preparation/audio_extract.py --filelist $filelist  --video_root $video_root --audio_root $audio_root

$filelist: a txt file containing names of videos. We provide the filelist of LRW test set as an example in the datalist directory.
$video_root: root directory of videos. In LRS2 dataset, $video_root should contains directories like "639XXX". In LRW dataset, $video_root should contains directories like "ABOUT".
$audio_root: root directory for saving wavforms
other optional arguments: please refer to audio_extract.py

To detect bounding boxes in videos and save it:

python preparation/bbx_extract.py --filelist $filelist  --video_root $video_root --bbx_root $bbx_root --gpu $gpu

$bbx_root: a root directory for saving detected bounding boxes
$gpu: run bbx_extract on a specific gpu. For example, 3.

*If you want to accelerate bbx_extract via multi-thread processing, you can use the following bash script:

*Please revise variables in the 2-nd to the 9-th lines to make it compatible with your own machine.

sh preprocess.sh

$file_list_dir: a directory which contains train.txt, valid.txt, test.txt of LRS2 dataset
$num_thread: number of threads you used. Please do not let it cross 8 with a 24GB GPU, 4 with a 12GB gpu.

Checkpoints

Model	Description	Link
TalkLip (g)	TalkLip net with the global audio encoder	Link
TalkLip (g+c)	TalkLip net with the global audio encoder and contrastive learning	Link
Lip reading observer 1	AV-hubert (large) fine-tuned on LRS2	Link
Lip reading observer 2	Conformer lip-reading network	Link
Lip reading expert	lip-reading network for training of talking face generation	Link
Discriminator	Discriminator of GAN	Link

Train

Some AV-Hubert files need to be modified.

rm xxx/av_hubert/avhubert/hubert_asr.py
cp avhubert_modification/hubert_asr_wav2lip.py xxx/av_hubert/avhubert/hubert_asr.py

rm xxx/av_hubert/fairseq/fairseq/criterions/label_smoothed_cross_entropy.py
cp avhubert_modification/label_smoothed_cross_entropy_wav2lip.py xxx/av_hubert/fairseq/fairseq/criterions/label_smoothed_cross_entropy.py

You can train with the following command.

python train.py --file_dir $file_list_dir --video_root $video_root --audio_root $audio_root \
--bbx_root $bbx_root --word_root $word_root --avhubert_root $avhubert_root --avhubert_path $avhubert_path \
--checkpoint_dir $checkpoint_dir --log_name $log_name --cont_w $cont_w --lip_w $lip_w --perp_w $perp_w \
--gen_checkpoint_path $gen_checkpoint_path --disc_checkpoint_path $disc_checkpoint_path

$file_list_dir: a directory which contains train.txt, valid.txt, test.txt of LRS2 dataset
$word_root: root directory of text annotation. Normally, it should be equal to $video_root, as LRS2 dataset puts a video file ".mp4" and its corresponding text file ".txt" in the same directory.
$avhubert_root: path of root of avhubert (should like xxx/av_hubert)
$avhubert_path: download the above Lip reading expert and enter its path
$checkpoint_dir: a directory to save checkpoint of talklip
$log_name: name of log file
$cont_w: weight of contrastive learning loss (default: 1e-3)
$lip_w: weight of lip reading loss (default: 1e-5)
$perp_w: weight of perceptual loss (default: 0.07)
$gen_checkpoint_path(optional): enter the path of a generator checkpoint if you want to resume training from a checkpoint
$disc_checkpoint_path(optional): enter the path of a discriminator checkpoint if you want to resume training from a checkpoint

Note: Sometimes, discriminator losses may diverge during training (close to 100). Please stop the training and resume it with a reliable checkpoint.

Test

The below command is to synthesize videos for quantitative evaluation in our paper.

python inf_test.py --filelist $filelist --video_root $video_root --audio_root $audio_root \
--bbx_root $bbx_root --save_root $syn_video_root --ckpt_path $talklip_ckpt --avhubert_root $avhubert_root

$filelist: a txt file containing names of all test files, e.g. xxx/mvlrs_v1/test.txt.
$syn_video_root: root directory for saving synthesized videos
$talklip_ckpt: a trained checkpoint of TalkLip net

Demo

I update the inf_demo.py on 4/April as I previously suppose that the height and width of output videos are the same when I set cv2.VideoWriter(). Please ensure the sampling rate of the input audio file is 16000 hz.

If you want to reenact the lip movement of a video with a different speech, you can use the following command.

python inf_demo.py --video_path $video_file --wav_path $audio_file --ckpt_path $talklip_ckpt --avhubert_root $avhubert_root

$video_file: a video file (end with .mp4)
$audio_file: an audio file (end with .wav)

**Please ensure that the input audio only has one channel

Evaluation

Please follow README.md in the evaluation directory

Citation

@inproceedings{wang2023seeing,
  title={Seeing What You Said: Talking Face Generation Guided by a Lip Reading Expert},
  author={Wang, Jiadong and Qian, Xinyuan and Zhang, Malu and Tan, Robby T and Li, Haizhou},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={14653--14662},
  year={2023}
}

DmitriyVahrushev/TalkLip