Seed3D 2.0 generates simulation-ready 3D assets with up to 89.9% win rate

Seed3D 2.0, presented by a team of researchers, builds on Seed3D 1.0 with major upgrades in fidelity, simulation readiness, and application coverage. For geometry, a coarse-to-fine two-stage pipeline decouples global structure learning from high-frequency detail recovery, while a locality-aware VAE achieves higher spatial compression and more efficient decoding. For materials and textures, the system replaces the previous cascaded pipeline with a unified PBR (physically based rendering) model that directly generates multi-view albedo and metallic-roughness maps. This model is enhanced by Mixture-of-Experts (MoE) scaling and vision-language model (VLM) based semantic conditioning, boosting material precision and visual fidelity.

Beyond single-object generation, Seed3D 2.0 introduces a simulation-ready model suite that includes scene layout planning, part-aware decomposition, and training-free articulation generation. This enables coherent scene construction and part-level physical interaction across physics and graphics engines. A large-scale human preference study against five recent commercial models showed consistent win rates ranging from 69.0% to 89.9%, demonstrating strong user preference for Seed3D 2.0's textured 3D assets. The system is available via an official page and technical report on arXiv.