Quality-Centric Single-Image Procedural Material Generation

Abstract

Procedural materials, represented as functional node graphs, are ubiquitous in computer graphics for photorealistic material appearance design. They allow users to perform intuitive and precise editing to achieve desired visual appearances. However, even for experienced artists, creating a procedural material given an input image requires professional knowledge and significant effort. Current inverse procedural material modeling approaches enable the automatic generation of procedural materials from input images. However, the visual quality of the generated materials is fundamentally limited by insufficient high-quality training data from industry-standard procedural materials, reliance on token-space supervision without visual feedback, and the absence of approximation-free node parameter post-optimization. My thesis presents advanced dataset augmentation, model training, and parameter post-optimization algorithms to address these challenges, significantly improving the perceptual match between the generated procedural material and the input image. Furthermore, the methodologies can be applied to other inverse procedural graphics problems to expedite similar artistic creation processes.