SceneMat: Procedural Material Authoring for Multi-Object 3D Scenes via Reference-Guided Optimization
Abstract
Procedural materials, typically represented as computational graphs, are widely adopted in 3D content creation because they enable editable, reusable, and highly controllable appearance design. To automate this expertise-intensive and time-consuming process, we study the problem of reference-guided stylized procedural material authoring for multi-object 3D scenes. This task poses three main challenges: (1) existing methods mostly assume single-object material optimization, which is insufficient for multi-object scenes; (2) stylized procedural materials are strongly coupled with object semantics and geometry; (3) procedural material authoring requires joint optimization over both structure and parameters, resulting in a highly complex optimization space. To address these challenges, we propose SceneMat, a reference-guided framework for procedural material authoring. Given a text or image reference, SceneMat first performs scene analysis to extract semantic, geometric, and spatial cues; it then refines scene-level references into object-level cues and performs material initialization; finally, it conducts material optimization through restricted VLM-driven iterative editing based on a structured intermediate representation. Experimental results show that the proposed method outperforms existing baselines in both image-guided and text-guided settings, achieving superior object-level and scene-level visual quality, as well as higher optimization correctness and efficiency.