Investigation of the structure of the latent variable space in Sentence-BERT sentence vectors using an image generation model

抄録

In this paper, we explore the intricate relationship between colors within the latent variable space, employing an image generation model tailored to produce images in response to textual prompts structured as “color + car type + car name”. Upon leveraging this model to analyze the interrelation of colors across various car types, it was observed that color positions within the latent variable space showcased remarkable consistency across car categories, evoking the semblance of an equilateral triangle. Moreover, for specific car types, three RGB color points were identified as emblematic and were utilized to construct a plane. By modulating the latent variables within this established plane, it was discerned that while the car's shape remained invariant, the color underwent modifications. This leads us to postulate the potential existence of hyperplanes in the latent variable space, each emblematic of distinct car classifications. This investigation seeks to elucidate the interplay between color and shape information encapsulated within the latent variable space.