SCHEMA: A Task-Decoupled Generative Framework for Procedural Materials via Cognitive Semantic Alignment
Abstract
Procedural materials provide editable, resolution-independent representations essential for virtual asset authoring. Yet existing methods mainly address appearance reconstruction or text-guided generation, where physical attributes, perceptual cues, and subjective impressions are entangled in a single description, causing ambiguous generation targets and limiting semantic control over executable procedural graphs. To address this gap, we propose SCHEMA (Semantic-Cognitive Hierarchical Engine for Material Authoring), a two-stage framework that separates semantic grounding from structural recovery. We support this decomposition with a four-level description framework spanning physical, perceptual, impressionistic, and emotional dimensions, and use it to construct a validated dataset of 529,558 image-description-code triplets, comprising 12,424 curated base triplets and 517,134 purified augmented triplets. A diffusion-based aligner first maps hierarchical descriptions to a rendered material appearance that preserves material structure while serving as a unified visual anchor, and a vision-language generator then translates this appearance into executable Blender procedural code. On the VLMaterial Blender test split, SCHEMA improves CLIP similarity by ~8.6% and DINOv2 structural similarity by ~20% over VLMaterial, while achieving 97.87% code executability on out-of-distribution data. In two blinded pairwise studies (n = 27 and n = 33), participants favored the structured hierarchy and the full pipeline on affective match and overall quality under the tested conditions.