README.md

One-Shot Free-View Neural Talking Head Synthesis

Unofficial pytorch implementation of paper "One-Shot Free-View Neural Talking-Head Synthesis for Video Conferencing".

Python 3.6 and Pytorch 1.7 are used.

Updates:

2021.11.05 :

• Replace Jacobian with the rotation matrix (Assuming J = R) to avoid estimating Jacobian.
• Correct the rotation matrix.

2021.11.17 :

• Better Generator, better performance (models and checkpoints have been released).

Driving | Beta Version | FOMM | New Version:

driving-beta-fomm-new.mp4

---

Driving | FOMM | Ours:

Free-View:

Train:

python run.py --config config/vox-256.yaml --device_ids 0,1,2,3,4,5,6,7

Demo:

python demo.py --config config/vox-256.yaml --checkpoint path/to/checkpoint --source_image path/to/source --driving_video path/to/driving --relative --adapt_scale --find_best_frame

free-view (e.g. yaw=20, pitch=roll=0):

python demo.py --config config/vox-256.yaml --checkpoint path/to/checkpoint --source_image path/to/source --driving_video path/to/driving --relative --adapt_scale --find_best_frame --free_view --yaw 20 --pitch 0 --roll 0

Note: run crop-video.py --inp driving_video.mp4 first to get the cropping suggestion and crop the raw video.

Pretrained Model:

Model	Train Set	Baidu Netdisk	Media Fire
Vox-256-Beta	VoxCeleb-v1	Baidu (PW: c0tc)	MF
Vox-256-New	VoxCeleb-v1	-	MF
Vox-512	VoxCeleb-v2	soon	soon

Note:

1. For now, the Beta Version is not well tuned.
2. For free-view synthesis, it is recommended that Yaw, Pitch and Roll are within ±45°, ±20° and ±20° respectively.
3. Face Restoration algorithms (GPEN) can be used for post-processing to significantly improve the resolution.

Acknowlegement:

Thanks to NV, AliaksandrSiarohin and DeepHeadPose.