MatComposer: Structural-Semantic Retrieval and Iterative Exploration for Procedural Material Graph Authoring
Abstract
Procedural materials are essential for resolution-independent surface appearance design, yet authoring remains expertise-intensive because users must translate high-level visual intent into complex node-based logic. While generative models can automate synthesis, they often produce static rasterized outputs that bypass the structured, editable representations required for professional iteration and asset reuse. We present MatComposer, an interactive system for authoring editable procedural material graphs through retrieval, exploration, and composition. MatComposer introduces a slot-based representation that decomposes material descriptions into functional modules, supported by intent parsing with Large Language Models (LLMs) and a dual-space retrieval pipeline that jointly considers semantic relevance and structural suitability for recomposition. The system enables users to perform slot-wise iterative refinement from intermediate previews and automatically composes selected sub-graphs into functional, editable procedural graphs. A within-subjects user study (N = 20) shows that, for non-expert users, MatComposer improves exploration efficiency, follow-up editing efficiency, and perceived editability over baseline retrieval workflows.