LGM 3D

What is LGM 3D?

LGM 3D (Large Multi-View Gaussian Model) is a framework for generating high-resolution 3D content from text prompts or a single image. It represents 3D objects using multi-view Gaussian features, which are efficient to render, and processes multiple views with an asymmetric U-Net backbone designed for high throughput. The result is high-fidelity 3D objects generated in under 5 seconds, with a training resolution of 512 — a balance of speed and quality.

Who created LGM 3D?

LGM 3D was developed by a group of researchers: Jiaxiang Tang and Xiaokang Chen from Peking University, Zhaoxi Chen and Ziwei Liu from S-Lab at Nanyang Technological University, and Tengfei Wang from Shanghai AI Lab.

How it works

1. Input: You provide a text prompt or a single-view image.
2. Multi-view generation: The model generates multiple views of the object from your input, often using multi-view diffusion models.
3. Representation: Multi-view Gaussian features represent the 3D object and are combined for rendering.
4. Backbone: An asymmetric U-Net processes the multi-view images quickly, enabling generation in about 5 seconds at a training resolution of 512.

This pipeline makes LGM 3D a fast option for producing detailed, high-resolution 3D models from minimal input.