Latent Space Uniformization in Generation

Jiali Song, Xuebin Chang, jianghao su, Bicong Ren, Wei Zeng

Keywords: latent space representation, generation, alignment, transformation

Abstract: Latent space transformation is a core topic in generative model research and is crucial for understanding and controlling the generative process. This study proposes the concept of latent space uniformization inspired by Coulomb's law, named ULatent. This concept provides a canonical representation of the latent space and facilitates sampling and aligning elements in both single-domain and cross-domain generative scenarios. Specifically, we model data points in a two-dimensional latent space as charged particles driven by Coulomb forces (electrostatic dynamics). The repulsive forces between them form a uniform distribution in the latent space, simulating the phenomenon where equally charged particles reach equilibrium. The uniformization of the original data enhances latent space structure, particularly by eliminating gaps between isolated clusters. For semantically overlapping clusters, pre-translation operations are required. By integrating geometric mapping techniques, we achieve precise alignment of uniformly distributed data across both single-modal and multi-modal domains, thereby simultaneously improving sampling efficiency and generation accuracy. All these conclusions are validated through multi-dataset experiments and ablation studies.

Primary Area: generative models

Submission Number: 11428