* Equal technical contribution.
A simulator needs more from an object than its shape. It needs to know which parts are which, how stiff each one is, how they hold together under contact — and a mesh lifted from a photograph carries none of that reliably.
DiagGen generates the asset and then checks its own work. Part-aware geometry and material properties come from an everyday image; the result is then put through controlled interactions in a physics simulator, and a diagnostic agent probes the regions where the outcome is most informative. What it finds is routed back as repair cues — to segmentation, to material inference, or to mesh processing, depending on where the fault lies.
Evaluated on 40 assets, the feedback is useful and repair gives moderate improvements. The generated assets hold up in contact-rich manipulation and drop into reconstructed real-world scenes, where their connectivity and material behaviour make a visible difference downstream.
The asset gallery, interactive 3D previews and the full method are on the project site.