Mode collapse is when a generator learns to produce only a few outputs.
It happens because those particular outputs reliably fool the discriminator.
The samples look fine one at a time, but the variety has gone.
It is the classic failure mode of training a GAN.
Diffusion models suffer from it far less, which is part of why they took over.
Mode collapse is when a generator learns to produce only a narrow set of outputs.
A mode is a peak in the data distribution, such as one style of face.
Real data has many, and a good generator should cover all of them.
A collapsed generator has found a few that reliably fool the discriminator and stopped exploring.
The samples look fine one at a time, which is what makes it easy to miss.
Look at a hundred and you see the same handful of things.
Trained on handwritten digits, it might produce only threes and eights.
The cause is in the objective, since the generator is only rewarded for fooling the judge.
Nothing in that reward asks for variety.
If one output works, producing it every time is a winning strategy.
The discriminator should catch the repetition and it only sees one sample at a time.
Minibatch discrimination lets it see several at once, which is one fix.
Unrolled GANs and Wasserstein loss attack the same problem differently.
Diffusion models suffer far less, since they are trained to reconstruct real data rather than to fool a judge.
That is a large part of why they took over image generation.
One question on this card. Every option explains itself once you pick, right or wrong.
18 concepts in this module, then a graded test with every answer explained.