Exploring Diffusion Models’ Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks

During few-shot fine-tuning, image quality can improve, collapse into noisy patterns, and then recover with severe overfitting. We model this corruption stage as a narrowed learned distribution and mitigate it with Bayesian neural networks.

Recommended citation: Wu, X.*, Zhang, J.*, Hua, Y., et al. (2026). "Exploring Diffusion Models' Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks." KDD 2026.
Download Paper