mode collapse
Mode collapse is the classic GAN failure where the generator, instead of reproducing the full variety of the data, gets stuck producing only a few kinds of output — or even one. Imagine a counterfeiter who discovers that one particular fake reliably fools the detective, so it just prints that one over and over. The images may look individually realistic, but the generator has collapsed the rich data distribution down to a few "modes", losing diversity.
A "mode" is a high-probability region of the data distribution (for example, each digit class is a mode of MNIST). Mode collapse occurs because the standard generator objective rewards fooling the current discriminator, not covering the data. If mapping many different z to the same convincing output fools D, the generator is happy to do so — nothing in the per-sample loss demands diversity. The alternating minimax optimization aggravates this: the generator can chase whichever mode the discriminator currently undervalues, the discriminator adapts, the generator hops to another mode, producing oscillation (a cat-and-mouse cycle) rather than coverage. This connects to the mode-seeking asymmetry of the Jensen–Shannon / reverse-KL-flavored objective.
Many fixes attack it. Wasserstein GAN with gradient penalty gives gradients that better penalize missing mass, improving coverage. Minibatch discrimination / minibatch standard deviation lets the discriminator see diversity within a batch, so all-identical outputs are flagged. Unrolled GANs let the generator anticipate the discriminator's future response, discouraging mode-hopping. PacGAN feeds the discriminator packed multiple samples at once. Other levers include experience replay of past fakes, the two-time-scale update rule (TTUR), spectral normalization, and conditional GANs (conditioning on labels forces per-class coverage). Diversity is then measured with metrics like recall (versus precision) or the number of distinct modes recovered on synthetic data.
Train a GAN on MNIST and watch the samples: in a healthy run you see all ten digits in varied styles; under mode collapse the generator might emit only crisp 1s and 7s regardless of the input noise z — every z decodes to nearly the same handful of images.
Do not confuse mode collapse with mere blur or low quality. Collapsed samples can be sharp and individually perfect; the defect is between samples — too little variety. That is why single-image quality scores miss it and you need diversity-sensitive metrics (recall, coverage, or FID, which jointly penalizes fidelity and diversity).