I've been trying to train a GAN using an image set of 2D video game sprites with a solid color background of a random color.
There's been a peculiar issue where the output of my generator is images with exclusively two colors. Specifically these two colors have RGB values that are either 1 or 0. The two colors are constant across every epoch. If I restart with a new generator it might be a new set of two colors, but after the first epoch, the colors remain constant.
Images from two different epochs of the same generator:
My generator is fairly basic, composed of several 2D transpose convolutional with batch normalization layers between them, and a few regular 2D convolutional layers, ending with a tanh activation. Because the generator uses a tanh activation at the end, the outputs of the layers before the tanh are relatively positive or negative. I've altered the exact order and makeup of the layers on a variety of instances, but results remain the same.
My discriminator is just a set of convolution layers with batch normalization between them, and I've tried using both Cross Entropy Loss and Wasserstein Loss. Again the exact makeup of the discriminator has varied. Both the generator and discriminator use Leaky ReLU between the layers.
It's honestly just kind of perplexing as to why this is happening in the first place.
Difficult to say without seeing more of your implementation. But generally this would imply that your model is collapsing. This is most likely due to one of your hyperparameters.
Gans are difficult to train and hyperparameter tuning can have a big effect on your output.
I would suggest simplifying your network and adjusting the learning rate to see how that effects the output and experiment from there.