THESIS DEFENSE - Beichen Li: Quality-Centric Single-Image Procedural Material Generation

Procedural materials, represented as functional node graphs of texture generation and image processing operators, are ubiquitous in modern computer graphics production 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 to visually match an input image requires professional knowledge and significant effort. Current inverse procedural material modeling approaches employ multi-modal Transformers to automatically generate procedural materials from single flash photos of object surfaces. However, the generated materials are fundamentally limited in visual quality due to: 1) insufficient high-quality training data from artist-created materials; 2) a lack of visual feedback in token-space supervised training; 3) the absence of approximation-free node parameter post-optimization for noise/pattern generator nodes. My thesis proposes advanced dataset augmentation techniques, training methodologies, and parameter post-optimization workflows to address these challenges, significantly improving the perceptual match between the generated procedural material and the input image. Furthermore, the research ideas are applicable to other inverse design problems in procedural graphics to expedite similar artistic creation processes.