What's this about?
Agents watch a video and write a graphics program that reproduces not only appearance but movement, deformation, fluid flow or fracture.
What exactly is inverse graphics?
Normal graphics turns a scene description into pixels. Inverse graphics works backward from pixels to the objects, rules and parameters that could have produced them.
Why does this matter?
Executable reconstruction tests causal understanding more strongly than visual similarity. Current frontier models remain much less reliable on dynamics than on static scenes.