Neural Appearance Synthesis and Transfer

Abstract

Appearance acquisition is a challenging problem. Existing approaches require expensive hardware and acquisition times are long. Alternative ''in-the-wild'' few-shot approaches provide a limited reconstruction quality. Furthermore, there is a fundamental tradeoff between spatial resolution and the physical sample dimensions that can be captured in one measurement. In this paper, we investigate how neural texture synthesis and neural style transfer approaches can be applied to generate new materials with high spatial resolution from high quality SVBRDF measurements. We perform our experiments on a new database of measured SVBRDFs.