LRM: Large Reconstruction Model for Single Image to 3D

• DOI: 10.48550/arxiv.2311.04400
• Publication Date: 2023-01-01
• Authors: Yicong Hong; Kai Zhang; Jiuxiang Gu; Sai Bi; Yang Zhou; Difan Liu; Feng Liu; Kalyan Sunkavalli; Trung Bui; Hao Tan
• Keywords: Image stitching; 3d model; Generative model

We propose the first Large Reconstruction Model (LRM) that predicts 3D model of an object from a single input image within just 5 seconds. In contrast to many previous methods are trained on small-scale datasets such as ShapeNet in category-specific fashion, LRM adopts highly scalable transformer-based architecture with 500 million learnable parameters directly predict neural radiance field (NeRF) image. train our end-to-end manner massive multi-view data containing around 1 objects, including both synthetic renderings Objaverse and real captures MVImgNet. This combination high-capacity large-scale training empowers be generalizable produce high-quality reconstructions various testing inputs, real-world in-the-wild images created by generative models. Video demos interactable meshes can found project webpage: https://yiconghong.me/LRM.