Generating Interior Images with Latent User Preferences Through GANs

The emergence of vast amounts and variety of data on the internet has driven the development of systems that gen-erate images tailored to users' preferences. However, extracting and understanding these preferences remains challenging due to their inherent complexity and ambiguity. In this paper, we present a method for modeling user preferences and generating images that align with those preferences. The proposed method employs Generative Adversarial Networks (GANs) to learn the latent space of user preference information, which is then estimated using Markov Chain Monte Carlo (MCMC) techniques. Specifically, the method utilizes reRDE-MC, a replica exchange Monte Carlo method capable of estimating multimodal distributions, which has been shown to effectively capture the complexities of user preferences. To address the reduced sampling accuracy of reRDE-MC in high-dimensional spaces like GAN latent spaces, we group generated images to facilitate the identification of user preferences. The effectiveness of the proposed method is validated by assuming a preferred image in the latent space and generating an image that aligns with the user's preference. The results confirm that reRDE-MC effectively extracts user preferences, unlike conventional MCMC methods that can only capture a single preference.