Creating a 3D object that you can interact with, like opening a lid or bending a part, from just one photo is hard. The photo shows how the object looks but not how it moves or what parts are hidden. Existing methods often produce a static shape or a moving model that no longer looks like the original object.
PhysAg3D uses a vision-language agent that first guesses the object’s parts and how they should move, while a standard 3D generator makes a detailed mesh of the whole object. A special representation links these guesses to the actual surfaces, so the agent can add joints or flexible parts while keeping the original appearance.
Tests on real photos show the generated assets look like the object and behave plausibly, such as a compass lid closing or a telescope reorienting. This matters for populating virtual worlds or robot simulations with realistic, interactive objects from a single picture.