Hallucination-Aware Diffusion Sampling for Inverse Problems via Robust Prior Updates

2026-06-01

AI Summary

Diffusion models often generate hallucinated content inconsistent with observed measurements when solving inverse problems. This work identifies the prior update stage as the primary source of such hallucinations and introduces Robust Prior Update (RPU)—a plug-and-play solver-level module that suppresses hallucinations through local stability analysis and a displacement re-anchoring strategy, without altering the measurement conditioning step. Experiments demonstrate that RPU significantly improves PSNR and LPIPS on FFHQ, achieves 91.9% preference in blind human evaluations and 91.1% under ground-truth-assisted assessment, and shows consistent advantages on ImageNet. This study is the first to explicitly attribute hallucinations to the prior update mechanism, thereby enhancing the fidelity of reconstructions to the original measurements.

Abstract

Diffusion-based inverse problem solvers can produce realistic reconstructions, but realism alone does not ensure that the recovered details are supported by the measurement. We study this failure as measurement-conditioned hallucination: visually meaningful content that is either implausible or inconsistent with the measured instance. Our analysis separates Bayes-rule-based diffusion inverse solvers into a prior update and a measurement-conditioning step, showing that hallucinated content can enter through the prior-side proposal before the measurement correction is applied. Motivated by this view, we propose Robust Prior Update (RPU), a solver-level module that probes the local stability of the diffusion prior update, re-anchors the resulting displacement at the current iterate, and leaves the measurement update unchanged. We instantiate RPU in DPS and evaluate it on FFHQ and ImageNet inverse problems using automatic metrics and human faithfulness studies. On FFHQ, RPU improves PSNR and LPIPS over DPS across box inpainting, Gaussian deblurring, and motion deblurring. In human judgments, RPU receives 91.9% of blind non-tie majority preferences and 91.1% of ground-truth-assisted non-tie preferences on FFHQ box inpainting, while the ImageNet Gaussian reader study is tie-heavy but favors RPU among non-tie cases. These results support a targeted claim: robustifying the prior update can improve instance faithfulness in diffusion inverse solvers, especially when the prior shapes weakly constrained content.

Problem

• hallucination
• diffusion models
• inverse problems
• measurement consistency
• prior update

Innovation

• diffusion models
• inverse problems
• hallucination mitigation
• robust prior update
• measurement consistency