Do generative models learn rare generative factors?

Generative models are becoming a promising tool in AI alongside discriminative learning. Several models have been proposed to learn in an unsupervised fashion the corresponding generative factors, namely the latent variables critical for capturing the full spectrum of data variability. Diffusion Models (DMs), Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) are of particular interest due to their impressive ability to generate highly realistic data. Through a systematic empirical study, this paper delves into the intricate challenge of how DMs, GANs and VAEs internalize and replicate rare generative factors. Our findings reveal a pronounced tendency toward memorization of these factors. We study the reasons for this memorization and demonstrate that strategies such as spectral decoupling can mitigate this issue to a certain extent.