I am still new to GANs and trying to implement the One to Many CGAN, on the MNIST Dataset, to generate a sequence of images (with the number of output images equal to the number of generators) from a sum. For example, A 4-generators 1-discriminator CGANS model will generate 4 digit number images: [3, 2, 6] if the inputs are [noise, noise, noise, noise] and condition = 11.

However, I encountered a problem where the Discriminator’s loss kept decreasing to near zero, but the Generators’ increased. Therefore, the Generators produce noises instead of meaningful images.

I supposed this is due to the fact that the Discriminator was too good, so I added Dropout layers and decreased the number of filters but nothing changed.

I am trying to get this model to work properly as I described above.

Here is my code.