laoreja / face-identity-transformer Goto Github PK

Why is the FR loss less than 0 during the train?And why is the G loss also less than 0?It's Strange.

Hi, I tried your pre-trained model and observed that it works fine for the given batch size. But, whenever I make inferences with batch size 1 the model is not anonymizing the face and there are many dark spots on the generated faces at random places. Additionally, when I try passing the same password to all the images in a batch, then also the models fail to do the anonymization. I suspect the model is batch dependent and it has learned that way. What do you think about this?

Hello author,I has successfully downloaded the trained model, however the compression pack of checkpoint was found to be damaged .Could you send me a copy of the compression package,thanks a million! Here is my email:[email protected]
Thanks again!

Hello author, I am very interested in your work and I want to have a try, but there may be some error with the Google Drive Folder link of the trained model which cannot be downloaded. Could you please share it again?

Thanks!

Thanks for publishing the nice work.

In the configuration of the data folder, I wonder about where to download the FFHQ's 'validation_paths.txt'.
Is it okay to extract thumbnail-file paths (val) from FFHQ's JSON file?

I look forward to hearing from you.
Have a nice day.

Thanks