DarkGeekMS/portrait-synthesis-from-speech

Integrate generated feature directions to current text-to-face generation

Closed this issue · 2 comments

Integrate the generated feature directions to directions-based text-to-face generation pipeline. This includes :

  • Adjusting BERT text processing to generate the same order of features are feature axes.
  • Adjusting the latent manipulation to accommodate the new feature axes.

The latent manipulation process is modified to eliminate the usage of logits normalization and scale factor, which yields better results through scaling the text logits in a specific range for safe navigation.
These changes are included in 40200b3

Also, the integration process includes text logits post-processing, in order to have the same order of features for both text and feature directions.