Local Intrinsic Dimension Unveils Hallucinations in Diffusion Models

TL;DR

This paper introduces a new perspective on diffusion model hallucinations, identifying local intrinsic dimension as a key factor and proposing a corrective method called Intrinsic Quenching to improve structural consistency.

Contribution

It reveals local intrinsic dimension as the main driver of hallucinations and introduces Intrinsic Quenching, a novel method to reduce these hallucinations effectively.

Findings

Intrinsic Quenching outperforms existing hallucination reduction methods.

LID is identified as the primary driver of hallucinations.

IQ improves anatomical consistency in medical imaging tasks.

Abstract

Diffusion models are prone to generating structural hallucinations - samples that match the statistical properties of the training data yet defy underlying structural rules, resulting in anomalies like hands with more than five fingers. Recent research studied this failure mode from several viewpoints, offering partial explanations to their occurrence, such as mode interpolation. In this work, we propose a complementary perspective that treats hallucinations as instabilities on the model-induced manifold. We begin by showing that a hallucination filter based on such instabilities matches or exceeds the performance of the recently proposed temporal one. By tracing the source of these instabilities, we identify local intrinsic dimension (LID) as their primary driver and propose Intrinsic Quenching (IQ), a direct corrective mechanism that deflates it to alleviate hallucinations. IQ consistently outperforms standard hallucination reduction baselines across a wide array of benchmarks and offers a highly promising solution for enforcing anatomical consistency in downstream medical imaging tasks.