# Cross-Property Coupling

PhysVGGT formulates physical property estimation as dense per-pixel prediction: a visual geometry transformer extracts geometry-aware tokens, a dense branch predicts per-pixel maps of friction coefficient, Shore hardness, Young's modulus and density, and a global branch predicts object-level mass, all from a single RGB image in one forward pass, achieving state-of-the-art mass estimation on ABO-500 and state-of-the-art friction and hardness on the out-of-distribution NeRF2Physics benchmark with 0.13 s inference latency, 27 times faster than the previous state of the art.

PhysVGGT formulates physical property estimation as dense per-pixel prediction: a visual geometry transformer extracts geometry-aware tokens, a dense branch predicts per-pixel maps of friction coefficient, Shore hardness, Young's modulus and density, and a global branch predicts object-level mass, all from a single RGB image in one forward pass, achieving state-of-the-art mass estimation on ABO-500 and state-of-the-art friction and hardness on the out-of-distribution NeRF2Physics benchmark with 0.13 s inference latency, 27 times faster than the previous state of the art.