Pytorch-cyclegan-and-pix2pix: Hint for network flexibility

[shinigami1992]

Hi, thanks for your great work.

I'm trying to apply cyclegan on my own custom anime dataset to learn facial expression transfer.
My dataset includes 2000 sad and 2000 happy images.

I run the training process for about 100k iterations.

The problem is that I only see some small changes specially in mouth part. Do you have any suggestion? How can I make network more flexible for changing different parts of faces?

Some sample results

[junyanz]

I recommend that you use a smaller lambda for cycle consistency loss, remove identity loss, or increase the receptive field of the discriminator. You can also experiment with a bigger generator.