Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion

2025-02-28

AI Summary

Diffusion models are prone to memorizing training samples—especially on small datasets—while existing unlearning methods often degrade generation quality. This work identifies low-noise scales as the primary source of memorization and proposes Ambient Diffusion, a novel paradigm that trains DDPMs exclusively at high-noise scales. By initializing with large Gaussian noise and applying noise truncation during sampling, Ambient Diffusion decouples high-frequency detail reconstruction from semantic content generation, enabling joint optimization of fidelity and low memorization. The method is compatible with both text-conditional and unconditional generation and builds upon standard UNet architectures without architectural modifications. Experiments on CIFAR-10 show a 92% reduction in memorization rate, while maintaining or improving FID and CLIP-Score. Furthermore, Ambient Diffusion consistently outperforms baselines across multiple dataset sizes, demonstrating robustness and scalability.

Technology Category

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

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems

Abstract

There is strong empirical evidence that the state-of-the-art diffusion modeling paradigm leads to models that memorize the training set, especially when the training set is small. Prior methods to mitigate the memorization problem often lead to a decrease in image quality. Is it possible to obtain strong and creative generative models, i.e., models that achieve high generation quality and low memorization? Despite the current pessimistic landscape of results, we make significant progress in pushing the trade-off between fidelity and memorization. We first provide theoretical evidence that memorization in diffusion models is only necessary for denoising problems at low noise scales (usually used in generating high-frequency details). Using this theoretical insight, we propose a simple, principled method to train the diffusion models using noisy data at large noise scales. We show that our method significantly reduces memorization without decreasing the image quality, for both text-conditional and unconditional models and for a variety of data availability settings.

Problem

• Memorization in diffusion models affects generation quality.
• Trade-off between fidelity and memorization in generative models.
• Training diffusion models with noisy data reduces memorization.

Innovation

• Trains diffusion models with large noise scales
• Reduces memorization without quality loss
• Applies to text-conditional and unconditional models