Abstract:Learning the distribution of data on Riemannian manifolds is crucial for modeling data from non-Euclidean space, which is required by many applications in diverse scientific fields. Yet, existing generative models on manifolds suffer from expensive divergence computation or rely on approximations of heat kernel. These limitations restrict their applicability to simple geometries and hinder scalability to high dimensions. In this work, we introduce the Riemannian Diffusion Mixture, a principled framework for building a generative diffusion process on manifolds. Instead of following the denoising approach of previous diffusion models, we construct a diffusion process using a mixture of bridge processes derived on general manifolds without requiring heat kernel estimations. We develop a geometric understanding of the mixture process, deriving the drift as a weighted mean of tangent directions to the data points that guides the process toward the data distribution. We further propose a scalable training objective for learning the mixture process that readily applies to general manifolds. Our method achieves superior performance on diverse manifolds with dramatically reduced number of in-training simulation steps for general manifolds.

What problem does this paper attempt to address?

The paper aims to address the problem of learning data distributions in non-Euclidean spaces, particularly for data residing on Riemannian manifolds. Specifically, the paper tackles several key issues that existing generative models face when dealing with such data: 1. **Expensive divergence computation**: Existing generative models on manifolds often require complex divergence computations, which become very time-consuming and difficult to scale in high-dimensional spaces. 2. **Dependence on heat kernel estimation**: Some methods rely on approximate estimation of the heat kernel, which can lead to performance degradation. 3. **Lack of geometric intuition**: Previous parameterization methods for diffusion models (based on score functions) lack clear geometric interpretation. To address these issues, the paper proposes a new framework called "Riemannian Diffusion Mixture," which constructs the generative diffusion process based on a mixture bridge process. This approach does not require heat kernel estimation and effectively guides the process towards the target data distribution by designing the drift of the diffusion process as a weighted average in the tangent direction to the data points. The main contributions of the paper can be summarized as follows: - Proposes a principled framework for constructing generative processes on general manifolds, based on a mixture bridge process, without the need for heat kernel estimation. - Designs the drift of the diffusion process as a weighted average in the tangent direction on the manifold, guiding the process towards the target distribution through this geometric design. - Introduces an efficient training objective—bidirectional bridge matching, which is applicable to general manifolds and does not require divergence computation. - Experiments demonstrate that the proposed method performs superiorly on various manifolds and significantly accelerates training speed compared to previous diffusion models, especially when handling high-dimensional data. In short, this study proposes a new generative model framework that can efficiently handle data residing on complex manifolds and demonstrates its effectiveness and superiority both theoretically and experimentally.

Generative Modeling on Manifolds Through Mixture of Riemannian Diffusion Processes

Scaling Riemannian Diffusion Models

A Mixed-Curvature Graph Diffusion Model

Graph Generation with Diffusion Mixture

Learning nonlinear manifolds based on mixtures of localized linear manifolds under a self-organizing framework

Varying Manifolds in Diffusion: From Time-varying Geometries to Visual Saliency

Unified framework for diffusion generative models in SO(3): applications in computer vision and astrophysics

Isometric Representation Learning for Disentangled Latent Space of Diffusion Models

Manifold Diffusion Fields

Learning Distributions on Manifolds with Free-form Flows

Where to Diffuse, How to Diffuse, and How to Get Back: Automated Learning for Multivariate Diffusions

Latent Schr{ö}dinger Bridge Diffusion Model for Generative Learning

Data Interpolations in Deep Generative Models under Non-Simply-Connected Manifold Topology

Unraveling the Smoothness Properties of Diffusion Models: A Gaussian Mixture Perspective

A Flexible Diffusion Model

Manifold Diffusion Geometry: Curvature, Tangent Spaces, and Dimension

Do Diffusion Models Learn Semantically Meaningful and Efficient Representations?

Convergence of Diffusion Models Under the Manifold Hypothesis in High-Dimensions

Trans-Dimensional Generative Modeling via Jump Diffusion Models

Learning Divergence Fields for Shift-Robust Graph Representations

Trivialized Momentum Facilitates Diffusion Generative Modeling on Lie Groups