baldFemale
commented
5 years ago One big difference between the original paper and my code is that I simply use face parsing tool in dlib which cannot provides an accurate face mask but includes other parts such as hair or background. It explains part of unnatural face performance. The paper applies more detailed face parsing algorithm.
Addition to this problem, super parameters do matter. Because the original paper does not reveal many details, I can only change the weights experimentally based on my own normalization for each loss. Decreasing the weight of face loss and lip loss as well as increasing the cycle consistency loss leads to relatively better performance.
Theoretically spectral norm stables the training process but I haven't trained the model without spectral normalization. Instance norm should also work.
One more interesting finding. Non-makeup imgs showed in the original paper always have more appealing results than others.
vincentwei0919
huynx8888
Hi, I really appreciate your excellent work! I've tried to train the model for my personal use with all settings not changed. But generally I found that when the training finished, most of the original images can be recovered, but the fake images especially for those from non-makeup to makeup ones always not as good as your paper had pasted. I check the super parameters used for your code again and found that the weight face loss(contained in makeup loss) is not 0.1 but equally used with lip/eye loss, so is this may be the reason? Or if your spectral norm matters? what if I stopped using this kind of norm? I've tried to change the weights for each loss, but not so useful, if there is a principal or range for these weights? Or just experimental values? Waiting for your reply! Thank you again!