1. Non-convergence
minmax objective cycle generated so repeated. Even with a small learning rate, it will not converge
example.
2. Mode-Collapse
What if the generator keeps generating a single realistic image? The discriminator will be always fooled by the single sample
Mini Batch Trick
Compute the similarity of the image with other images in the same batch to avoid Mode-Collapse. This measures the diversity of the batch.
Feed the similarity score along with the image to the discriminator as an input feature. This penalizes the generator and encourages it to generate less similar images
Many more advanced techniques have also been proposed so far.