search

theEricMa / DiffSpeaker

This is the official repository for DiffSpeaker: Speech-Driven 3D Facial Animation with Diffusion Transformer
137 stars 18 forks source link

About # of transformerdecoder layers #10

Closed samsara-ku closed 5 months ago

samsara-ku commented 5 months ago

Is it okay to set just num_layers as 1 for TransformerDecoderLayer_w_Adapter class when training your model?

Some other models just use their num_layers as 6 or 8, is this just a simple curiosity.

theEricMa commented 5 months ago

We follow the configuration in FaceFormer, which uses just one layer of Transformer. You can add more layers, which can improve performance.