Laplacian Score Sharpening for Mitigating Hallucination in Diffusion Models

The recent submission of 'Laplacian Score Sharpening for Mitigating Hallucination in Diffusion Models' on arXiv addresses a critical issue in AI, where diffusion models generate unrealistic samples due to hallucinations. These hallucinations arise from mode interpolation and score smoothening, which have been inadequately managed in prior research. The authors propose a novel post-hoc adjustment to the score function during inference, leveraging the Laplacian of the score to mitigate these hallucinations. Their methodology includes deriving an efficient Laplacian approximation for higher dimensions using a finite-difference variant of the Hutchinson trace estimator. The results indicate a significant reduction in hallucinated samples across both toy 1D/2D distributions and high-dimensional image datasets. This work not only enhances the fidelity of AI-generated content but also explores the relationship between the Laplacian and uncertainty in the score, paving the way for more reliabl…