search

ParthaEth / GIF

GIF is a photorealistic generative face model with explicit 3D geometric and photometric control.
https://gif.is.tue.mpg.de/
MIT License
405 stars 63 forks source link

train on my own datasets ,but the test results are bad #18

Closed visonpon closed 3 years ago

visonpon commented 3 years ago

Hi, @ParthaEth, I have followed your step to create my own datasets, but after train for a few days, the test results are all bad(I only use one person(6k images) to train, and it contain abundant expression and pose )

329292484

I first use DECA to get parameters and then create the datasets use the scripts in prepare_lmdb My loss values are nearly like below: G_loss=5.x/ D_loss=1.x/embd_reg_loss=0/interp_loss=2.x

Hope you can give some advice, thanks~

ParthaEth commented 3 years ago

If you had good FID during train, My best guess is that at test time your rendered meshes are not normalized properly, i.e. same as train time. Different normalization? Just put a debug point ans see the range of values.