Title: VSM: Mitigating Hallucinations in Diffusion Models via Score Control

Abstract: Diffusion models have become the foundational architecture for contemporary generative AI, driving progress across vision, language, audio, and other domains. However, their widespread adoption is hampered by hallucinations—generated samples that are implausible and fall outside the actual data distribution—which undermine user trust and system reliability. This study first provides empirical validation for the hypothesis that score smoothness is a primary driver of hallucinations in image generation diffusion models, offering a new perspective rooted in data density. We further formalize this relationship by connecting the probability mass of hallucinations to the Lipschitz constant of the learned score function. Building on these insights, we propose Variance-Guided Score Modulation (VSM), a novel strategy that regulates the score Jacobian to reduce score smoothness. This adjustment allows for a more accurate approximation of the ground truth score, thereby suppressing hallucinations. Our experiments on both synthetic and real-world datasets show that VSM can reduce hallucination rates by up to ~25% without compromising sample fidelity or diversity, marking a significant advance toward more dependable diffusion-based image synthesis. Additionally, we introduce two new benchmark datasets characterized by extreme semantic variation to facilitate systematic evaluation of hallucinations. The associated code and datasets are accessible at https://github.com/bhosalems/VSM.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC