Mode collapse occurs when a GAN's generator maps all (or most) noise vectors z to a narrow region of the output space, ignoring most modes of p_data. The discriminator can be fooled locally, but global coverage fails. Root cause: the JS divergence can be zero even when distributions don't overlap, giving G no gradient signal toward missing modes. Solutions include Wasserstein loss (provides gradient everywhere), minibatch discrimination (D sees batch statistics), unrolled GANs (G trains against future D steps), and feature matching (G matches intermediate D statistics). The left panel shows true data distribution; the right shows generator samples with increasing collapse.