LGM (Large Gaussian Model)

LGM, the Large Multi-View Gaussian Model from 3DTopia, generates high resolution 3D assets from a text prompt or a single image in a matter of seconds. Presented as an ECCV 2024 oral paper, it first uses multi-view diffusion models, MVDream for text input and ImageDream for images, to synthesize consistent views of the target object, then feeds those views to an asymmetric U-Net that directly predicts a 3D Gaussian splatting representation. Inference requires roughly 10 GB of GPU memory with all components loaded, and fp16 variants reduce the footprint. Results are saved as PLY files of Gaussians, can be inspected in a local GUI, rendered to MP4 turntable animations, and converted to standard GLB meshes for use in game engines and 3D software. The official implementation is released under the MIT license, serving researchers working on feed forward 3D generation and developers experimenting with rapid 3D asset creation.