Understanding and Mitigating Copying in Diffusion Models
This addresses a critical privacy and copyright issue for users and developers of text-to-image diffusion models, offering incremental improvements to mitigate data copying.
The paper tackles the problem of diffusion models replicating training data, finding that text conditioning significantly contributes to memorization, and proposes techniques like caption randomization and augmentation to reduce data replication.
Images generated by diffusion models like Stable Diffusion are increasingly widespread. Recent works and even lawsuits have shown that these models are prone to replicating their training data, unbeknownst to the user. In this paper, we first analyze this memorization problem in text-to-image diffusion models. While it is widely believed that duplicated images in the training set are responsible for content replication at inference time, we observe that the text conditioning of the model plays a similarly important role. In fact, we see in our experiments that data replication often does not happen for unconditional models, while it is common in the text-conditional case. Motivated by our findings, we then propose several techniques for reducing data replication at both training and inference time by randomizing and augmenting image captions in the training set.