BANG: Dividing 3D Assets via Generative Exploded Dynamics (arxiv.org)

BANG introduces "Generative Exploded Dynamics," a new generative pipeline that decomposes 3D assets into coherent parts by producing a smooth sequence of exploded states that progressively separate geometry while preserving semantic and geometric consistency. Technically, the system fine-tunes a pre-trained large-scale latent diffusion model with a lightweight exploded-view adapter and a temporal attention module to enforce smooth transitions across the sequence. Users can steer decomposition with spatial prompts (bounding boxes, surface regions) and extend control via multimodal LLMs like GPT-4 for 2D-to-3D manipulations, letting designers specify or sketch which components to split and how. For the AI/ML and graphics communities, BANG bridges generation and reasoning to make part-level modeling far more intuitive and automatable. It produces detailed part-level geometry, links parts to functional descriptions, and can output separable components suited for 3D printing or manufacturing workflows. Beyond accelerating asset creation and lowering the artistic barrier, BANG has implications for dataset generation (automatic part segmentation), robotics and CAD-aware design, and interactive multimodal tooling that combines diffusion models with LLM-guided intent.