Redirecting to original paper in 30 seconds...

Click below to go immediately or wait for automatic redirect

arxiv_cv 95% Match Research Paper AI Researchers,Machine Learning Engineers,Computer Vision Practitioners 2 weeks ago

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

generative-ai › diffusion
📄 Abstract

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.
Authors (6)
Yiqi Tian
Pengfei Jin
Mingze Yuan
Na Li
Bo Zeng
Quanzheng Li
Submitted
July 16, 2025
arXiv Category
cs.CV
arXiv PDF

Key Contributions

Introduces RODS, a novel diffusion sampling method inspired by robust optimization, which detects and corrects high-risk sampling steps using geometric cues from the loss landscape. This method reduces hallucinations without retraining and at minimal inference cost, improving sampling fidelity and robustness.

Business Value

Enhances the reliability and quality of generated images, which is crucial for applications in creative industries, synthetic data generation, and content creation.