CarveAnything: Image-to-volumetric material with destruction mechanics

CarveAnything takes an input monocular image and generates a 3d model with predictive volumetric content, physics properties, and destruction mechanics for simulating cutting, lighting, and staging objects--so it may be adapted to the Halloween season.

One can use this to automate robotics policy evaluation and training on diverse objects, and more on methods is in the paper. We use a combination of semantic features and predictive per-Gaussian physics to transform individual complex objects or entire scenes to simulation environments in Isaac Lab or MuJoCo. Pumpkin carving itself is a difficult task to model accurately in these simulators.

Happy carving, HN! Glad to know of failure modes with any generated objects/carves.