Saved in:
Bibliographic Details
Main Authors: Tian, Yiqi, Jin, Pengfei, Yuan, Mingze, Li, Na, Zeng, Bo, Li, Quanzheng
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2507.12201
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914105646907392
author Tian, Yiqi
Jin, Pengfei
Yuan, Mingze
Li, Na
Zeng, Bo
Li, Quanzheng
author_facet Tian, Yiqi
Jin, Pengfei
Yuan, Mingze
Li, Na
Zeng, Bo
Li, Quanzheng
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2507_12201
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle RODS: Robust Optimization Inspired Diffusion Sampling for Detecting and Reducing Hallucination in Generative Models
Tian, Yiqi
Jin, Pengfei
Yuan, Mingze
Li, Na
Zeng, Bo
Li, Quanzheng
Computer Vision and Pattern Recognition
Optimization and Control
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.
title RODS: Robust Optimization Inspired Diffusion Sampling for Detecting and Reducing Hallucination in Generative Models
topic Computer Vision and Pattern Recognition
Optimization and Control
url https://arxiv.org/abs/2507.12201