Reproduce “Conffusion: Confidence Intervals for Diffusion Models” on OpenResearch

About this paper

Eliahu Horwitz, Yedid Hoshen

Researchers at The Hebrew University of Jerusalem developed CONffusion, a method that equips pre-trained diffusion models with statistically rigorous pixel-wise confidence intervals. This approach provides tight uncertainty bounds while achieving a three-orders-of-magnitude speedup over sampling-based methods, addressing a critical limitation for deploying diffusion models in high-stakes applications.

Key insights

• Leveraging the powerful U-Net backbone of pre-trained diffusion models allows for predicting confidence bounds in a single forward pass, leading to significant computational efficiency gains.
• Diffusion pre-training is crucial for generating tighter and more accurate confidence intervals compared to training from scratch or using discriminative models, due to the rich distributional knowledge encoded in these models.
• The Risk-Controlling Prediction Sets (RCPS) framework provides robust, distribution-free statistical guarantees, ensuring that the predicted intervals maintain a desired coverage probability.

Summary by alphaXiv.