Meta 3D AssetGen

Meta 3D AssetGen, published at NeurIPS 2024, tackles the problem most 3D generators quietly ignore: baked-in lighting. Instead of textures with shadows painted on, AssetGen decomposes the output into albedo, metalness and roughness — real PBR materials — so the generated mesh can be relit in a new environment and still look right. The pipeline generates a few views of the object in factored shaded and albedo channels, reconstructs shape via a signed-distance function with direct shape supervision, then sharpens textures with a refinement transformer in UV space.

The paper reports a 17% improvement in Chamfer Distance and 40% in LPIPS over the best concurrent few-view reconstruction work, plus a 72% human preference rate over comparable-speed industry competitors including PBR-capable ones. Those are author-reported numbers from the paper's own evaluation, not an independent benchmark. It is a research publication, not a product: no public weights, no API, no hosted demo. The project page shows generated assets and the paper is freely available.

Paper-reported: 17% better Chamfer Distance, 40% better LPIPS vs concurrent work; 72% human preference over comparable-speed industry competitors (author-reported)