Trellis
Trellis creates high-quality 3D assets from simple text or image inputs. Using a unified latent space (SLAT), it delivers detailed, textured 3D models in formats like meshes,radiance fields, and 3D Gaussians. Its flexibility, editing capabilities, and superior quality enable faster, more adaptable workflows in gaming, virtual worlds, industrial design, and beyond. Go deeper here.
Open Trellis Repo
Trellis was released for research purposes. In September 2025, NVIDIA added Trellis to its AI Blueprints solution.
Trellis provides a powerful foundation for scalable, AI-driven content creation. Built to meet rapidly growing industrial demand, Trellis creates high-quality, editable 3D assets from simple text or image prompts in lieu of manual modeling. This saves time, lowers barriers, and unlocks new possibilities for developers, designers, and digital content creators.
Trellis is built on a novel Structured LATent (SLat) representation that captures both geometric structure and visual detail in a compact, editable form. Trained on 500,000 diverse 3D objects using rectified flow transformers with up to 2 billion parameters, Trellis significantly improves both quality and flexibility compared to existing models.
Unlike traditional methods that target a single output type or require labor-intensive setup, Trellis can generate a 3D asset in multiple formats, including meshes, 3D gaussians, and radiance fields. This makes it compatible with different rendering pipelines and applications. The generated models feature detailed structure and rich texture, enabling their direct use in games, AR/VR experiences, digital twins, simulation environments, and product visualization.
Trellis also allows for prompt-guided local edits—such as removing, replacing, or adding parts of a 3D model—without retraining or manual sculpting, which dramatically accelerates iteration and customization. Its design eliminates the need for costly 3D fitting and leverages pretrained vision models for high-fidelity results, even when working with sparse 3D data.