TRELLIS.2 - Microsoft's open-source 4 billion parameter 3D generative model

What is TRELLIS.2?

TRELLIS.2 is an open-source 4 billion-parameter 3D generative model from Microsoft, designed for the efficient generation of high-fidelity 3D assets from single images. Its core innovation lies in its O-Voxel sparse voxel architecture, capable of handling complex topologies (such as open surfaces and non-manifold geometry) and supporting full PBR material modeling. It generates 512³ resolution models in just 3 seconds on an NVIDIA H100, and supports ultra-high-definition rendering up to 1536³. Compared to traditional methods, it significantly improves efficiency through 16x spatial downsampling, while also providing GLB format export functionality, making it suitable for games, film, and industrial design.

Main functions of TRELLIS.2

• Image to 3D generationIt can quickly generate high-quality 3D models from input images, making it suitable for scenarios where 3D assets are created from 2D images.
• Text to 3D generationIt supports generating 3D models based on text descriptions, allowing users to generate the desired 3D objects using text commands.
• Complex topology processingIt can handle complex 3D structures, including open surfaces, non-manifold geometry, and internally closed structures, generating 3D models with complex shapes.
• Material and Texture GenerationThe generated 3D model supports a variety of material properties (such as base color, roughness, metallicity, opacity, etc.) and supports transparency and realistic rendering.
• Multi-resolution supportIt supports the generation of 3D models at various resolutions, from 512³ to 1536³, allowing users to choose the appropriate resolution according to their needs.
• Fast rendering and exportThe generated 3D models can be used directly for rendering and can be exported to standard 3D file formats such as GLB, making them easy to use in other software.
• Model Editing and Variant GenerationUsers can edit the generated 3D model to create different variations of the same object, meeting diverse design needs.

TRELLIS.2 Technical Principles

• O-Voxel indicatesIt adopts the "O-Voxel" sparse voxel structure, which is an innovative 3D representation method that can efficiently handle complex topologies and sharp features, breaks through the limitations of traditional isosurface fields, and supports open surfaces, non-manifold geometry and internal closed structures.
• Sparse 3D VAE encodingIt encodes 3D assets into a compact latent space through a sparse 3D variational autoencoder (VAE), achieving efficient compression and generation, supporting 16x spatial downsampling, and significantly improving processing efficiency.
• Full texture modelingIt not only processes basic colors, but also simulates PBR material properties such as roughness, metallicity, and opacity, supports transparency and realistic rendering, and generates 3D models with rich texture details.
• Simplified data processing workflowData processing requires no complex rendering or optimization steps, and the conversion between texture mesh and O-Voxel can be completed in a short time, which greatly simplifies the processing flow and improves the generation speed.
• High-efficiency rendering and post-processingIt combines CUDA-accelerated mesh processing tools (such as CuMesh) with efficient rendering technology to quickly generate high-quality 3D models and supports real-time rendering and post-processing operations.
• Large-scale pre-trainingPre-trained on a large-scale dataset, the model learns rich 3D shape and texture features, enabling it to generate diverse 3D assets based on input images or text.

TRELLIS.2 project address

• Project official websitehttps://microsoft.github.io/TRELLIS.2/
• Github repository: https://github.com/microsoft/TRELLIS.2
• HuggingFace model libraryhttps://huggingface.co/microsoft/TRELLIS.2-4B
• arXiv technical paper: https://arxiv.org/pdf/2512.14692
• Experience the demo onlinehttps://huggingface.co/spaces/microsoft/TRELLIS.2

Application scenarios of TRELLIS.2

• Game developmentIt can quickly generate high-quality 3D game assets, such as characters, props, and scenes, improving development efficiency and reducing production costs.
• Film and television productionIt is used to create realistic 3D effects and virtual scenes, supports complex materials and textures, and meets the high requirements of the film and television industry for visual effects.
• Virtual Reality (VR) and Augmented Reality (AR)It generates immersive 3D environments and objects, enhancing the user experience and is suitable for VR/AR application development.
• Architectural Design and VisualizationIt can quickly generate 3D visualizations of architectural models and interior designs, helping designers and clients better understand and evaluate design options.
• Industrial Design and Product DevelopmentCreate 3D models of product prototypes to accelerate the design iteration process and facilitate virtual demonstrations and functional testing.
• Education and TrainingAs a teaching tool, it helps students and trainees better understand and master the relevant knowledge and skills of 3D modeling, rendering and design.