What type of study is this?

This is a Quantitative Study study (also classified as: Experimental Study).

October 16, 2025Open Access

Improving the Euclidean Diffusion Generation of Manifold Data by Mitigating Score Function Singularity

View Full Paper

Ask AI

Bookmark

Authors

ZLZichen LiuPeng Cheng Laboratory WZWei ZhangUniversity of Hawaiʻi at Mānoa TLTiejun LiXi'an University of Science and Technology

Discussion

Member takes

Implication

This analysis reveals score function singularities in manifold data, suggesting improved sampling methods. It introduces Niso-DM and Tango-DM to bolster accuracy in generative models.

Key Points

The study uncovers multiscale singularity in score functions, impacting the accuracy of diffusion-generated samples.
Niso-DM uses non-isotropic noise to improve performance across manifold distributions, significantly enhancing sampling accuracy.
Tango-DM employs a tangential-only loss function, refining the training of the score function's tangential component for better modeling.
Numerical experiments validate the proposed methods' superior performance with complex geometric distributions over manifolds.

Cite This Study

Liu et al. (2025) studied this question.

synapsesocial.com/papers/68f147cc724575985c3fd225 https://doi.org/10.48550/arxiv.2505.09922

View Full Paper

Ask AI

Bookmark

Implication

This analysis reveals score function singularities in manifold data, suggesting improved sampling methods. It introduces Niso-DM and Tango-DM to bolster accuracy in generative models.

Improving the Euclidean Diffusion Generation of Manifold Data by Mitigating Score Function Singularity

Discussion

Member takes

Implication

Key Points

Cite This Study

Improving the Euclidean Diffusion Generation of Manifold Data by Mitigating Score Function Singularity

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider