search

universome / epigraf

[NeurIPS 2022] Official pytorch implementation of EpiGRAF
https://universome.github.io/epigraf
150 stars 6 forks source link

Custom FFHQ Training #9

Closed zhukaii closed 1 year ago

zhukaii commented 1 year ago

When the custom ffhq dataset is adopted without camera pose information, the error exists in train.py (line 127, "Broken yaw angles (all zeros)"). So how can I reproduce the ffhq training result similar to the one in paper?

universome commented 1 year ago

Hi! Camera conditioning is a crucial ingredient to learn decent geometry for FFHQ since it contains too many frontal images and the learnt geometry will be flat otherwise. So, I do not think that it is possible to reproduce the results without using camera poses from GRAM (which you can download from their repo).