GRM: Large Gaussian Reconstruction Model for Efficient 3D Reconstruction and Generation

We introduce GRM, a large-scale reconstructor capable of recovering a 3Dasset from sparse-view images in around 0.1s. GRM is a feed-forwardtransformer-based model that efficiently incorporates multi-view information totranslate the input pixels into pixel-aligned Gaussians, which are unprojectedto create a set of densely distributed 3D Gaussians representing a scene.Together, our transformer architecture and the use of 3D Gaussians unlock ascalable and efficient reconstruction framework. Extensive experimental resultsdemonstrate the superiority of our method over alternatives regarding bothreconstruction quality and efficiency. We also showcase the potential of GRM ingenerative tasks, i.e., text-to-3D and image-to-3D, by integrating it withexisting multi-view diffusion models. Our project website is at:https://justimyhxu.github.io/projects/grm/.