how to decrease identity loss, the final result looks too much like 70newface-30%oldface mix no not like swap to fully new face

[2blackbar]

[momstouch]

see line 218 in train.py you can control id-loss using opt.lambda_id

[Fibonacci134]

I believe he means with the current model, in which case not much can be done. you can play with the masking and the swapped img a bit if youre familiar with opencv.