Latent Space Map for Visual Utilization of Generated Data.

Abstract

Samples produced by generative models, called Generated Samples (GSs), have become a critical supplement to those collected from the real world in data-centric applications. Domain experts typically randomly collect many GSs and manually select a few of interest for applications. However, the methodology lacks guidance to locate desirable ones that exhibit specific features or adhere to application-oriented metrics among infinite generable candidates. These samples are generally concentrated in a few small regions of the generative model's latent space, called Generative Latent Space (GLS). This paper presents Latent Space Map that projects a GLS onto a plane to help users locate regions rich in desirable GSs. Our research revolves around two challenges in constructing the map. First, many GSs in a GLS are low-quality and useless for applications. Excluding them from the projection is challenging for their irregular distribution. We employ a Monte Carlo-based method to capture a manifold for projection, where high-quality GSs are mainly distributed. Second, the GLS is high-dimensional and unbounded, complicating the projection. We design a manifold projection method that endows the map with desirable characteristics to achieve high display accuracy and effective pattern perception for users freely observing the manifold. We further develop a system integrating Latent Space Map to aid in GS selection and refinement. Real-world cases, quantitative experiments, and feedback from domain experts confirm the usability and effectiveness of our approach.