Generalization of Diffusion Models Arises with a Balanced Representation Space

Diffusion models generate high-quality, diverse images with great generalizability, yet when overfit to the training objective, they may memorize training samples. We analyze memorization and generalization of diffusion models through the lens of representation learning. Using a two-layer ReLU denoising autoencoder (DAE) parameterization, we show that memorization corresponds to the model learning the raw data matrix for encoding and decoding, yielding spiky representations; in contrast, generalization arises when the model captures local data statistics, producing balanced representations. We validate these insights by investigating representation spaces in real-world unconditional and text-to-image diffusion models, where the same distinctions emerge. Practically, we propose a representation-based memorization detection method and a simple representation-steering method that enables controllable editing of generalized samples. Together, our results underscore that learning good representations is central to novel and meaningful generation.