Using Deep Generative Models for Glossy Appearance Synthesis and Exploration

Generating images with realistic material appearance using a physically-based renderer demands significant time and human labor. The images are used in psychophysical experiments to study human perception of material appearance attributes, such as glossiness. Recently, deep learning-based image synthesis models have emerged as a promising approach for generating realistic images with less human supervision. Deep Generative Models are deep learning-based models that learn to generate unique and novel images based on a given training data distribution. Using them for image synthesis is fast and manually less tiresome. An additional benefit these Deep Generative Models offer is latent space encodings that may help to better understand the feature space of gloss and its perception. In this study, we propose to explore the possibility of using Deep Generative Models for realistic image synthesis, focusing on gloss appearance and evaluating the efficiency of such gloss generation process using psychophysical experiments. Additionally, we build tools to extract the latent space of generative models to use them as a feature space representation of gloss appearance and perception. Finally, we analyse the trends and patterns in the learnt feature space to aid gloss appearance modelling.