Mode Collapse
Mode collapse occurs when a generative model produces only a limited diversity of outputs, ignoring large parts of the data distribution.
Mode collapse = generative models produce only few variants instead of full diversity – the classic GAN problem that diffusion models have largely solved.
Explanation
In GANs, the generator finds a "safe" strategy that fools the discriminator and keeps producing similar images. Diagnostics: FID/IS metrics, visual inspection of samples, nearest-neighbor analysis against training data.
Marketing Relevance
Mode collapse is the main problem in GAN-based content generation – monotonous outputs are useless for marketing.
Example
A GAN for product images always generates the same angle and background, despite diverse training data.
Common Pitfalls
Noticing mode collapse only late in training. Only checking FID score (can be good despite collapse). Diffusion models have this problem much less frequently.
Origin & History
Mode collapse was recognized early as a GAN problem (Goodfellow, 2014). Wasserstein GAN (Arjovsky, 2017) and Spectral Normalization (Miyato, 2018) reduced the problem. Diffusion models (2020+) largely solved it through likelihood-based training.
Comparisons & Differences
Mode Collapse vs. Overfitting
Overfitting copies training data exactly; mode collapse ignores parts of the distribution and generates only certain patterns.
Mode Collapse vs. Posterior Collapse (VAE)
Mode collapse in GANs: generator ignores modes. Posterior collapse in VAEs: encoder ignores inputs and uses only the prior.