TRELLIS.2: state-of-the-art large 3D generative model (4B) (github.com)

TRELLIS.2, a cutting-edge large 3D generative model featuring 4 billion parameters, has been unveiled, significantly enhancing the capabilities of image-to-3D asset generation. This model utilizes a novel "field-free" sparse voxel structure known as O-Voxel, which enables the reconstruction and generation of complex 3D assets while preserving intricate details, such as textures and multiple surface attributes including roughness and opacity. With resolutions ranging from 512³ to 1536³, TRELLIS.2 offers exceptional rendering speed—approximately 3 seconds for 512³ output on an NVIDIA H100 GPU, demonstrating its efficiency and scalability for a variety of applications in virtual environments and gaming. The significance of TRELLIS.2 lies in its ability to handle a broader range of geometries, including open and enclosed structures, without degradation in quality. This offers a transformative potential for industries requiring high-fidelity 3D models, such as gaming, film, and virtual reality. The streamlined data processing and integration capabilities mean users can quickly convert and optimize 3D assets while leveraging features like photorealistic rendering and rapid mesh generation. With pre-trained models available on platforms like Hugging Face and an upcoming release of a demo tool, TRELLIS.2 is poised to advance the landscape of AI-driven 3D content creation by simplifying workflows and enhancing asset quality.