AssetGen: Interactive 3D Asset Generation in 30 Seconds
A novel AI model named AssetGen creates high-quality 3D meshes with baked normals and color textures from a single reference image in just 30 seconds, making it suitable for real-time rendering, including mobile applications. This system, developed by researchers and outlined in arXiv:2605.26137, employs a coarse-to-refine VecSet framework for geometry, along with GPU-based mesh simplification and normal baking, and rapid parallel UV unwrapping. Textures are generated using multi-view techniques, backprojection, and 3D inpainting. Additionally, a quicker version called AssetGen Flash brings latency down to 14 seconds for interactive and agentic creation processes. The model integrates model distillation, kernel optimization, and pipeline parallelization to enhance deployability at interactive speeds, focusing on user experience and deployability, which are often neglected in recent 3D generation studies.
Key facts
- AssetGen generates 3D assets from one reference image in 30 seconds.
- AssetGen Flash variant reduces latency to 14 seconds.
- Outputs include high-quality mesh with baked normals and color texture.
- Polygon budget is controlled for real-time rendering, including mobile.
- Uses coarse-to-refine VecSet framework for geometry generation.
- GPU implements mesh simplification, cleaning, and normal baking.
- Fast parallel UV unwrapping is employed.
- Textures generated via multi-view approach with backprojection and 3D inpainting.
Entities
Institutions
- arXiv