tool · draft
Extend3D
Research model from Seoul National University that generates a town-scale 3D scene from a single image, built on Microsoft's TRELLIS; code released with a CVPR 2026 paper.
What is reusable: open sourceunknown dataAPI unknownunknown participation MITCode on GitHub under MIT; the quick start loads pretrained TRELLIS weights.
Extend3D is a research model for town-scale 3D generation from Seungwoo Yoon, Jinmo Kim and Jaesik Park at Seoul National University, published at CVPR 2026. The code is built on Microsoft's TRELLIS and the repository credits that project.
The quick start takes one image and returns a scene as 3D Gaussians and a mesh. The repository shows rendering the Gaussians to a video and exporting the mesh as a GLB file. A Gradio app gives a browser demo.
It runs as installed Python code under the MIT licence. It needs a Linux x86-64 machine, an NVIDIA GPU with 24 GB of memory, and CUDA 12.4 or later; larger scenes need more memory. The quick start loads the pretrained weights named microsoft/TRELLIS-image-large.
Status: research code, released alongside the paper. The setup notes say "Coming Soon", and the paper and demo links in the README are not yet live. In this wiki it sits in the engines layer beside other 3D reconstruction and generation models such as /engines/vggt/.
- form
- research model
- group
- 3D cities and mapping
- homepage
- https://github.com/SNU-VGILab/Extend3D
- licence
- MIT
- cost
- free
Tags: 3d generation, town, single image, gaussian splatting, mesh, computer vision