eeskimez / emotalkingface

The code for the paper "Speech Driven Talking Face Generation from a Single Image and an Emotion Condition"
MIT License
161 stars 29 forks source link

Speech Driven Talking Face Generation from a Single Image and an Emotion Condition

[Paper link]

Each row shows 6 videos created using the same image and speech but with different emotion input. Columns represent anger, disgust, fear, happiness, neutral, and sadness, respectively.

screen-gif

Installation

pip install -r requirements.txt

It depends on the following packages:

The code is tested on Ubuntu 18.04 and OS X 10.15.2.

If you have troubles with D-Lib (especially for Windows), please use conda install instead of pip:

conda install -c conda-forge dlib

Download Data

You can download the data from this repo.

Convert Videos to 25 FPS

Run the following code:

python .\data_prep\convertFPS.py -i \raw_video_folder -o \output_folder

Prepare Data

python .\data_prep\prepare_data.py -i \25_fps_video_folder\ -o \output_folder --mode 1 --nw 1

Train

First pre-train the emotion discriminator:

python train.py -i /train_hdf5_folder/ -v /val_hdf5_folder/ -o ../models/mde/ --pre_train 1 --disc_emo 1 --lr_emo 1e-4

Then pre-train the generator:

python train.py -i /train_hdf5_folder/ -v /val_hdf5_folder/ -o ../models/pre_gen/ --lr_g 1e-4

Finally, train all together:

python train.py -i /train_hdf5_folder/ -v /val_hdf5_folder/ -o ../models/tface_emo/ -m ../models/pre_gen/ -mde ../models/mde/ --disc_frame 0.01 --disc_emo 0.001

By default, the Tensorboard log file is written to the output path. You can check the intermediate video results and loss values using Tensorboard.

Inference

Download our pretrained model (optional): Please put pre-trained model in the model folder.

Inference from an image and speech file:

python generate.py -im ./data/image_samples/img01.png -is ./data/speech_samples/speech01.wav -m ./model/ -o ./results/

Inference from processed dataset (h5py files):

python generate_all_emotions.py -ih /path/to/h5py/folder/ -m ./model/ -o ./results/

Inference from processed dataset (h5py files) - Mismatched emotions:

python generate_mismatched_emotions.py -ih /path/to/h5py/folder/ -m ./model/ -o ./results/

Acknowledgment

We thank the authors of the following repo.

Citation

@ARTICLE{seeskimezemotface,
    title={Speech Driven Talking Face Generation from a Single Image and an Emotion Condition},
    author={Eskimez, Sefik Emre and Zhang, You and Duan, Zhiyao},
    journal={arXiv preprint arXiv:2008.03592},
    year={2020}
}