A question about the experimental result

[Danee-wawawa]

Hello, I am using uvcgan2 to make image set A have image set B image style while maintaining the original image content.
However, I am currently encountering an issue.
The generated 'fake_b.png' result completely lost the content of image "real_a.png" and approximated a certain image in the B training set.
What is the reason for this and how can it be resolved? I need help. Thank you~

[usert5432]

Hi @Danee-wawawa,

It is hard to tell without knowing the details of the dataset. My speculation is that the following things can be involved:

1. Insufficient magnitude of the cycle-consistency loss. To check this -- you can examine reco_a image. If reco_a looks like real_a then everything is fine. Otherwise, you could try increasing the magnitude of the lambda_a and lambda_b hyperparameters.
2. "Unaligned" datasets. If the datasets are very different in terms of the positions of the objects, this may also contribute. For instance, if one tries to perform Male <-> Female translation, but Male faces are located in the top 1/2 of the image, while Female faces are located in the bottom 1/2 of the image, then the resulting translation will fail to match Male <-> Female faces and produce random outputs. If your dataset exhibits similar properties, then you would need to add some data augmentations to fix that.

[Danee-wawawa]

OK，thank you~I will try these.