LRM: learning to reconstruct an object from one image at scale

What changed

A model can learn likely three-dimensional structure from many examples, then propose a new object quickly. LRM combines a pretrained image encoder with a camera-conditioned transformer that predicts three feature planes. A small neural renderer reads these planes to produce views or support mesh extraction. Hong and colleagues train this image-to-3D system on a large synthetic-and-real multi-view corpus. The contribution is the scaled reconstruction architecture and its demonstrated object generalization, with separate credit for inherited image features, transformers and neural rendering.

Scientific scope

A model can learn likely three-dimensional structure from many examples, then propose a new object quickly. LRM combines a pretrained image encoder with a camera-conditioned transformer that predicts three feature planes. A small neural renderer reads these planes to produce views or support mesh extraction. Hong and colleagues train this image-to-3D system on a large synthetic-and-real multi-view corpus. The contribution is the scaled reconstruction architecture and its demonstrated object generalization, with separate credit for inherited image features, transformers and neural rendering. The input is a foreground-preprocessed object image, with assumed normalized camera settings when the true camera is unknown. Incorrect cameras distort geometry. Training uses posed views, substantial compute and inherited pretrained features; an approximately million-scale training collection does not mean a million independently verified physical shapes. Hidden detail is predicted from priors. Background scenes and view-dependent materials remain limitations. A few-second reconstruction figure refers to specified A100 stages and hardware, not universal end-to-end performance or a controlled victory over every alternative.

Evidence

The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The input is a foreground-preprocessed object image, with assumed normalized camera settings when the true camera is unknown. Incorrect cameras distort geometry. Training uses posed views, substantial compute and inherited pretrained features; an approximately million-scale training collection does not mean a million independently verified physical shapes. Hidden detail is predicted from priors. Background scenes and view-dependent materials remain limitations. A few-second reconstruction figure refers to specified A100 stages and hardware, not universal end-to-end performance or a controlled victory over every alternative.

Contributors

• The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.

• LRM: Large Reconstruction Model for Single Image to 3D

Provisional WRS assessment, edition 2026-10-23.9. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction