TRELLIS is a free, open-source AI-powered 3D asset generation model developed by Microsoft Research that enables users to create high-quality 3D objects from text descriptions or reference images. The platform uses a novel Structured Latent (SLAT) representation and Rectified Flow Transformers to generate detailed 3D geometry, realistic textures, and visually complex objects. TRELLIS supports multiple 3D output formats, including polygon meshes, 3D Gaussian Splatting, and Radiance Fields, allowing developers and designers to choose the representation best suited to their projects. The software includes advanced editing capabilities that enable users to modify specific regions of generated objects, create alternative variations, and refine 3D models using additional text or image prompts. TRELLIS provides pretrained AI models with up to 2 billion parameters, trained on a curated dataset of 500,000 diverse 3D assets, along with tools for custom training and model fine-tuning.
Features
- AI Text-to-3D Generation: Generate detailed 3D objects from natural language descriptions using pretrained text-to-3D models, with configurable prompts and generation settings.
- Image-to-3D Model Creation: Transform reference images into three-dimensional assets with detailed geometry and textures, including experimental support for conditioning generation on multiple images.
- Multiple 3D Output Formats: Generate and export assets as polygon meshes, 3D Gaussian Splats, and Radiance Fields, with support for extracting GLB models for use in compatible 3D applications.
- Localized 3D Editing and Variations: Modify specific regions of generated 3D objects, add or remove structural elements, and create alternative versions while preserving the overall appearance and geometry.
- Pretrained Models and Custom Training: Access pretrained AI models ranging from 342 million to 2 billion parameters, along with training scripts, data preparation tools, and the TRELLIS-500K dataset for developing customized 3D generation models.
- GPU-Accelerated Generation and Interactive Demos: Generate and preview 3D assets using NVIDIA GPU acceleration, Python inference pipelines, and Gradio-based interfaces for experimenting with text-to-3D and image-to-3D workflows.