How Diffusion Models Memorize

2025-09-29

AI Summary

Diffusion models exhibit training-data memorization, posing privacy and copyright risks—yet the underlying mechanism remains unclear. This work identifies that memorization stems from overestimation of training samples during early denoising steps, rather than conventional overfitting; this bias causes classifier-free guidance to amplify training-image components in noise predictions, suppress stochasticity, and steer denoising trajectories toward specific memorized samples. Leveraging latent-space dynamical modeling, intermediate latent decomposition, denoising-path visualization, and a novel theoretical scheduling-bias quantification framework, we demonstrate a near-perfect correlation (≈1.0) between early-stage overestimation magnitude and memorization strength. This is the first interpretable, quantifiable theoretical foundation for memorization in generative models—providing dual utility for rigorous privacy-risk assessment and principled mitigation strategy design.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: (Large) Language Models

Application Category

• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Despite their success in image generation, diffusion models can memorize training data, raising serious privacy and copyright concerns. Although prior work has sought to characterize, detect, and mitigate memorization, the fundamental question of why and how it occurs remains unresolved. In this paper, we revisit the diffusion and denoising process and analyze latent space dynamics to address the question: "How do diffusion models memorize?" We show that memorization is driven by the overestimation of training samples during early denoising, which reduces diversity, collapses denoising trajectories, and accelerates convergence toward the memorized image. Specifically: (i) memorization cannot be explained by overfitting alone, as training loss is larger under memorization due to classifier-free guidance amplifying predictions and inducing overestimation; (ii) memorized prompts inject training images into noise predictions, forcing latent trajectories to converge and steering denoising toward their paired samples; and (iii) a decomposition of intermediate latents reveals how initial randomness is quickly suppressed and replaced by memorized content, with deviations from the theoretical denoising schedule correlating almost perfectly with memorization severity. Together, these results identify early overestimation as the central underlying mechanism of memorization in diffusion models.

Problem

• Investigates why diffusion models memorize training data during generation
• Analyzes latent space dynamics and denoising process mechanisms
• Identifies early overestimation of samples as core memorization driver

Innovation

• Analyzing latent space dynamics in diffusion models
• Identifying early overestimation as memorization mechanism
• Decomposing intermediate latents to reveal content suppression