PARSE Framework Introduces Part-Level Spatial Modeling for Physically Consistent 3D Scene Generation

Researchers from ShanghaiTech University and Deemos Technology have introduced PARSE (Part-Aware Relational Spatial Modeling), a framework that addresses a longstanding limitation in 3D scene generation: existing representations such as linguistic prepositions or object-level scene graphs are too coarse to reliably describe how objects physically interact. PARSE centers on the Part-centric Assembly Graph (PAG), which encodes geometric relationships between specific object parts, and a Part-Aware Spatial Configuration Solver that translates those relationships into geometric constraints for assembling collision-free, physically valid layouts.

To support training and evaluation, the team released PARSE-10K, a dataset of 10,000 3D indoor scenes built from real-image layout priors and a curated part-annotated shape database. Fine-tuning Qwen3-VL on PARSE-10K produced stronger object-level layout reasoning and more accurate part-level relation understanding; using PAGs as structural priors in 3D generation models yielded scenes with markedly improved physical realism and structural complexity.