Adaptive Destruction Processes for Diffusion Samplers

2025-06-02

AI Summary

This work addresses the challenge of efficiently sampling from unnormalized densities under data-free conditions in few-step diffusion models. The proposed method introduces a learnable, adaptive corruption process, decoupling the generative and corruption processes into independently trainable Gaussian transition kernels—thereby relaxing both the continuous-time assumption and fixed-variance constraints. It is the first to jointly optimize both processes under strict discrete-step limitations. Leveraging discrete-time policy optimization and unconstrained Gaussian parameterization, the approach significantly accelerates convergence and improves sample fidelity. Empirically, it achieves superior FID and LPIPS scores over state-of-the-art few-step diffusion methods across multiple benchmarks. Furthermore, it generalizes successfully to conditional image generation in GAN latent spaces, demonstrating strong scalability and broad applicability.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

This paper explores the challenges and benefits of a trainable destruction process in diffusion samplers -- diffusion-based generative models trained to sample an unnormalised density without access to data samples. Contrary to the majority of work that views diffusion samplers as approximations to an underlying continuous-time model, we view diffusion models as discrete-time policies trained to produce samples in very few generation steps. We propose to trade some of the elegance of the underlying theory for flexibility in the definition of the generative and destruction policies. In particular, we decouple the generation and destruction variances, enabling both transition kernels to be learned as unconstrained Gaussian densities. We show that, when the number of steps is limited, training both generation and destruction processes results in faster convergence and improved sampling quality on various benchmarks. Through a robust ablation study, we investigate the design choices necessary to facilitate stable training. Finally, we show the scalability of our approach through experiments on GAN latent space sampling for conditional image generation.

Problem

• Improving diffusion samplers with trainable destruction processes
• Decoupling generation and destruction variances for flexibility
• Enhancing convergence and sampling quality in limited steps

Innovation

• Trainable destruction process in diffusion samplers
• Decoupled generation and destruction variances
• Unconstrained Gaussian densities for transition kernels