microsoft/trellis.2

TRELLIS.2 is a research project from Microsoft that turns a single photograph into a detailed three-dimensional model. You give it an image, and it outputs a 3D asset complete with textures, color, roughness, and transparency information that you could use in a game engine, 3D editor, or rendering software. The model is large, with 4 billion parameters. It uses a new internal representation called O-Voxel, which is a way of storing 3D shapes that can handle tricky geometry like thin surfaces, holes, and objects with things inside them. Standard approaches in this area often struggle with those cases. The generation process takes around 3 to 60 seconds depending on the resolution you want, running on high-end NVIDIA graphics hardware. In addition to creating 3D shapes from images, the system can also generate textures for an existing 3D shape you provide. The output can be exported as a GLB file, which is a standard format for 3D objects that many applications can open. The project is research code intended for experimentation. It runs on Linux and requires an NVIDIA GPU with at least 24GB of memory. Setup involves installing a collection of specialized packages through a provided script, which the readme notes can take a while. A pretrained 4-billion-parameter model is available on Hugging Face, and there is also a live demo on Hugging Face Spaces where you can try it without setting anything up locally. The code and model weights are released under the MIT license.