Diffusion Model Hallucinations Linked to Local Intrinsic Dimension

A new study from arXiv (2605.05026) identifies local intrinsic dimension (LID) as a primary driver of structural hallucinations in diffusion models. These hallucinations produce samples that match training data statistics but violate structural rules, such as hands with extra fingers. The researchers propose Intrinsic Quenching (IQ), a corrective mechanism that deflates LID to reduce hallucinations. IQ outperforms standard hallucination reduction methods. The work offers a complementary perspective to existing explanations like mode interpolation, treating hallucinations as instabilities on the model-induced manifold.