EvelynFan / FaceFormer

[CVPR 2022] FaceFormer: Speech-Driven 3D Facial Animation with Transformers
MIT License
790 stars 134 forks source link

Training result not good #76

Open Aiedails opened 1 year ago

Aiedails commented 1 year ago

Hello! I'm trying to train the model from scratch, and I found even after 100 epochs on vocaset using the parameters you give, the result is depressing. The faces rarely move, and the performance is far from your given checkpoint's. Do you use other training tricks not listed in your paper or did I miss anything?

xjtupanda commented 1 year ago

Is that so? I've tried to train on the vocaset and checked some rendering results following the instruction. The result seems fine to me though.