RODS: Robust Optimization Inspired Diffusion Sampling for Detecting and Reducing Hallucination in Generative Models

TL;DR

RODS introduces a novel diffusion sampling method inspired by optimization principles to detect and reduce hallucinations in generative models, enhancing robustness and fidelity without retraining.

Contribution

The paper proposes RODS, a new diffusion sampling technique that detects and corrects hallucinations using geometric cues, improving robustness without additional training.

Findings

Detects over 70% of hallucinated samples

Corrects more than 25% of hallucinations

Maintains image quality and diversity

Abstract

Diffusion models have achieved state-of-the-art performance in generative modeling, yet their sampling procedures remain vulnerable to hallucinations-often stemming from inaccuracies in score approximation. In this work, we reinterpret diffusion sampling through the lens of optimization and introduce RODS (Robust Optimization-inspired Diffusion Sampler), a novel method that detects and corrects high-risk sampling steps using geometric cues from the loss landscape. RODS enforces smoother sampling trajectories and adaptively adjusts perturbations, reducing hallucinations without retraining and at minimal additional inference cost. Experiments on AFHQv2, FFHQ, and 11k-hands demonstrate that RODS maintains comparable image quality and preserves generation diversity. More importantly, it improves both sampling fidelity and robustness, detecting over 70% of hallucinated samples and correcting more than 25%, all while avoiding the introduction of new artifacts. We release our code at https://github.com/Yiqi-Verna-Tian/RODS.