We introduce PhysMAS, a physics‑grounded framework for synthesizing compositional 4D Gaussians in heterogeneous multi‑part and multi‑object scenes. Existing pipelines couple 3D Gaussians with the Material Point Method (MPM) but typically collapse distinct parts into a single material state, which compromises physical realism. One‑shot predictions from large language models or vision‑language models struggle to bind materials to identified parts and cannot verify whether the resulting MPM configuration is executable. Score Distillation Sampling (SDS)‑based optimization requires repeated per‑scene score evaluations and gradient back‑propagation, leading to long runtimes and potentially unstable solutions.
PhysMAS adopts a multi‑agent architecture:
- An Object‑Part Scene Agent extracts persistent identities for objects and their parts from a motion prompt and four scene views.
- A Material Reasoning Agent generates part‑wise material profiles.
- Solver‑aware skills bind these identities and profiles to per‑particle MPM fields and execute all objects within a shared domain.
- The framework then screens forward‑simulation outcomes, retaining candidates that satisfy both semantic and physical constraints.
By avoiding per‑scene SDS back‑propagation, PhysMAS reduces computational cost. Extensive experiments show that, compared with recent SDS‑based physics‑driven 4D Gaussian baselines, PhysMAS achieves better semantic alignment and perceived physical plausibility while requiring less runtime.
Review