honkywonkyman
DF Vagrant
I'm currently at iteration 20 000 training an SAE model on the Google Colab. Overall the training is progressing pretty nicely, but I've noticed that the eyes of the decoded face never seem to look in the same direction as the dest, and I'm worried this will turn into an issue later on in the training if I don't fix it. I also saw this picture of a Robert Downey JR - Elon Musk model, where you can clearly see that Elon's eyes don't match Robert's even though the model is at 85k iterations and the conversion is convincing otherwise:
Is this whole thing an issue with SAE or are there just some settings that need to be changed or something? If it's not SAE, do any of you have any ideas or suggestions on how to solve it?
Here are my training settings:
== Model name: SAE
==
== Current iteration: 19643
==
== Model options:
== |== write_preview_history : True
== |== target_iter : 40000
== |== batch_size : 8
== |== sort_by_yaw : False
== |== random_flip : False
== |== resolution : 128
== |== face_type : f
== |== learn_mask : True
== |== optimizer_mode : 1
== |== archi : df
== |== ae_dims : 512
== |== e_ch_dims : 42
== |== d_ch_dims : 21
== |== multiscale_decoder : True
== |== ca_weights : False
== |== pixel_loss : False
== |== face_style_power : 10.0
== |== bg_style_power : 10.0
== |== apply_random_ct : False
== |== clipgrad : False
== Running on:
== |== [0 : Tesla T4]
=========================
Example of what I mean. The face structure looks right except for the eyes.
Is this whole thing an issue with SAE or are there just some settings that need to be changed or something? If it's not SAE, do any of you have any ideas or suggestions on how to solve it?
Here are my training settings:
== Model name: SAE
==
== Current iteration: 19643
==
== Model options:
== |== write_preview_history : True
== |== target_iter : 40000
== |== batch_size : 8
== |== sort_by_yaw : False
== |== random_flip : False
== |== resolution : 128
== |== face_type : f
== |== learn_mask : True
== |== optimizer_mode : 1
== |== archi : df
== |== ae_dims : 512
== |== e_ch_dims : 42
== |== d_ch_dims : 21
== |== multiscale_decoder : True
== |== ca_weights : False
== |== pixel_loss : False
== |== face_style_power : 10.0
== |== bg_style_power : 10.0
== |== apply_random_ct : False
== |== clipgrad : False
== Running on:
== |== [0 : Tesla T4]
=========================
Example of what I mean. The face structure looks right except for the eyes.