Reconstruction works alright but results are poor

[ashutoshbsathe]

Hi @vanhuyz,

Thanks for the implementation. I'm working on transforming gameplay of "Prince of Persia 1" to "Prince of Persia 2". I have collected 8000+ images of both the games. I've been training for 2400+ steps right now and the results are as shown below -

Is this a case of model collapse ? If so how should I overcome this issue ?

Thanks again.

[Edit : I see original paper mentioning identity loss for overcoming some issues in the training. I also see that you have identity loss branch on this repo, but your commit for adding identity loss does not include identity loss in optimizer, am I correct ? Where should I add identity loss exactly ?]

[gunpowder1473]

I add it to the generator loss which used to train two generators. Improved the results in my implementation Hope helpful