PARSE: Part-Aware Relational Spatial Modeling
Researchers have introduced PARSE, a novel framework designed to improve spatial intelligence in AI by modeling interactions at the part level of objects. This approach utilizes a Part-centric Assembly Graph (PAG) to encode geometric relations between object parts, enabling the creation of physically consistent and collision-free 3D scenes. A new dataset, PARSE-10K, comprising 10,000 3D indoor scenes with detailed part-level annotations, was developed to train and evaluate the framework. Fine-tuning the Qwen3-VL model on this dataset demonstrated enhanced object-level layout reasoning and part-level relation understanding, while integration into 3D generation models resulted in scenes with greater physical realism. AI
IMPACT Enhances AI's ability to generate physically realistic 3D scenes and understand spatial relationships.