junyanz / pytorch-CycleGAN-and-pix2pix

Image-to-Image Translation in PyTorch

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

CycleGAN is bypassing information to minimize the reconstruction error

jgciudad opened this issue · comments

Hi, the results I obtained in the facade dataset look quite strange.

To me, it looks like if the CycleGAN is keeping the information in the synthetic generated images, so then it can get a perfect reconstruction of the input images, and is completely forgetting about optimizing the discriminators error. Note that even though the reconstruction of the facade looks identical to the original image, the following picture is with data from the test set, so is not an overfitting issue.

Does anybody has a hint of what can be happening? I saw this answer from #1184, which I don't know if could be somehow related.

Thanks!

imagen
(the black and red arrows represent the 2 cycle-GAN directions)

This limitation is documented by a well-written article "CycleGAN, a Master of Steganography". I am wondering if you are using the default parameters for the facades results. We have obtained better results compared to yours.