search

DinoMan / speech-driven-animation

949 stars 289 forks source link

Test result is not good when using real images #26

Closed CodingMice closed 4 years ago

CodingMice commented 4 years ago

It is a very job! But when I test the model using the images of myself, I got bad results. it seems the model has remembered the faces used in training.

DinoMan commented 4 years ago

Yes the models that have been released are from datasets (GRID, TCD TIMIT, CREMA-D) that have only a few faces. It will work on unseen faces from the same dataset where lighting and background are similar but doesn't work well for "in-the-wild" examples. The CREMA-D model has the most faces so it will likely work best. We have a model trained on a large set of faces but this has not been released. If something changes and we decide to release this model I will post it here.