Towards Memorization-Free Diffusion Models

TL;DR

Diffusion models risk memorizing training data during inference, raising copyright and privacy concerns. The paper introduces Anti-Memorization Guidance (AMG), a unified framework combining despecification (G_spe), caption deduplication (G_dup), and dissimilarity (G_sim) with an automatic memorization detection mechanism and a parabolic threshold schedule to selectively activate guidance. AMG yields memorization-free outputs across unconditional, class-conditional, and text-conditional generations on DDPM-based and latent diffusion models, with minimal impact on image fidelity and text alignment measured by FID and CLIP. This approach enables safer deployment of large-scale diffusion models on datasets like LAION5B while preserving output utility and alignment to user prompts.

Abstract

Pretrained diffusion models and their outputs are widely accessible due to their exceptional capacity for synthesizing high-quality images and their open-source nature. The users, however, may face litigation risks owing to the models' tendency to memorize and regurgitate training data during inference. To address this, we introduce Anti-Memorization Guidance (AMG), a novel framework employing three targeted guidance strategies for the main causes of memorization: image and caption duplication, and highly specific user prompts. Consequently, AMG ensures memorization-free outputs while maintaining high image quality and text alignment, leveraging the synergy of its guidance methods, each indispensable in its own right. AMG also features an innovative automatic detection system for potential memorization during each step of inference process, allows selective application of guidance strategies, minimally interfering with the original sampling process to preserve output utility. We applied AMG to pretrained Denoising Diffusion Probabilistic Models (DDPM) and Stable Diffusion across various generation tasks. The results demonstrate that AMG is the first approach to successfully eradicates all instances of memorization with no or marginal impacts on image quality and text-alignment, as evidenced by FID and CLIP scores.

Figures (15)

• Figure 1: Stable Diffusion's capacity to memorize training data, manifested as pixel-level memorization (left) and object-level memorization (right). Our approach successfully guides pretrained diffusion models to produce memorization-free outputs.
• Figure 2: Geometric interpretation of different guidance methods and generations. The center O represents a scenario where the generated image is identical to the memorized training image. The distance of any point from O reflects its degree of dissimilarity to the memorized image. The surface of the sphere signifies the threshold that defines the presence of a memorization issue. The arrows represent different types of guidance strategies.
• Figure 3: Comparison of similarity scores throughout the inference process, with and without the application of $G_{dup}$.
• Figure 4: Applying AMG to iDDPM on CIFAR-10. Left: Class-conditional generation. Right: Unconditional generation.
• Figure 5: Example illustrating the detection of potential memorization at an early stage of reverse diffusion process, enabling AMG to influence the coarser structures in generated outputs.