Elucidating Representation Degradation Problem in Diffusion Model Training

2026-05-11

AI Summary

Diffusion models often suffer from representation degradation under high noise levels, leading to structural distortions, training instability, and reduced generation quality. This work proposes the Elucidated Representation Diffusion (ERD) framework, which formally characterizes the root cause of this degradation for the first time and introduces a dynamic optimization reallocation mechanism that adaptively adjusts training resources based on the recoverability of target representations—without requiring external supervision. By leveraging neural tangent kernel (NTK) spectral analysis and recoverability assessment, ERD effectively mitigates NTK spectrum attenuation and low-rank effects. Extensive experiments demonstrate that ERD significantly accelerates convergence and enhances generation quality across diverse diffusion backbones, confirming its generality and effectiveness.

Abstract

Diffusion models have achieved remarkable success, yet their training remains inefficient due to a severe optimization bottleneck, which we term Representation Degradation. As noise levels increase, the outputs of the trained model exhibit progressive structural distortion, which can destabilize training and impair generation quality. Our analysis suggests that this instability is driven by mismatched target recoverability, which is associated with Neural Tangent Kernel (NTK) spectral weakening and effective low-rank behavior. To address this, we propose Elucidated Representation Diffusion (ERD), a plug-and-play framework that dynamically reallocates optimization effort according to effective recoverability. By stabilizing representation learning without external supervision, ERD accelerates convergence and achieves strong empirical performance across diffusion backbones.

Problem

• Representation Degradation
• Diffusion Models
• Optimization Bottleneck
• Structural Distortion
• Training Instability

Innovation

• Representation Degradation
• Neural Tangent Kernel
• Diffusion Models
• Effective Recoverability
• ERD