Emergence of Distortions in High-Dimensional Guided Diffusion Models

2026-01-31

AI Summary

Classifier-free guidance (CFG) in diffusion models suffers from variance collapse, leading to a loss of sample diversity manifested as a mismatch between the generated and true conditional distributions. This work formalizes this distortion as a high-dimensional distributional mismatch problem and, leveraging high-dimensional Gaussian mixture models, exact score functions, and dynamical mean-field theory, reveals a phase transition in the effective potential governing CFG dynamics. The analysis uncovers an emergent mechanism wherein this phase transition scales exponentially with the number of classes. Building on these insights, the paper proposes a theoretically grounded scheduling strategy incorporating a negative guidance window, which substantially mitigates diversity loss while preserving class separability—even in regimes where the number of modes grows exponentially with dimensionality.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Classifier-free guidance (CFG) is the de facto standard for conditional sampling in diffusion models, yet it often leads to a loss of diversity in generated samples. We formalize this phenomenon as generative distortion, defined as the mismatch between the CFG-induced sampling distribution and the true conditional distribution. Considering Gaussian mixtures and their exact scores, and leveraging tools from statistical physics, we characterize the onset of distortion in a high-dimensional regime as a function of the number of classes. Our analysis reveals that distortions emerge through a phase transition in the effective potential governing the guided dynamics. In particular, our dynamical mean-field analysis shows that distortion persists when the number of modes grows exponentially with dimension, but vanishes in the sub-exponential regime. Consistent with prior finite-dimensional results, we further demonstrate that vanilla CFG shifts the mean and shrinks the variance of the conditional distribution. We show that standard CFG schedules are fundamentally incapable of preventing variance shrinkage. Finally, we propose a theoretically motivated guidance schedule featuring a negative-guidance window, which mitigates loss of diversity while preserving class separability.

Problem

• generative distortion
• classifier-free guidance
• diffusion models
• high-dimensional regime
• diversity loss

Innovation

• generative distortion
• classifier-free guidance
• phase transition
• high-dimensional analysis
• negative-guidance schedule