OpenLRM

OpenLRM is an open source reimplementation of the Large Reconstruction Model, a feed-forward approach that turns a single image into a 3D representation in one pass rather than through slow per-shape optimization. Given one photo, it predicts a triplane NeRF and can export rendered videos and PLY mesh files. Released by 3DTopia, it comes in six pretrained variants across small, base, and large sizes, ranging from 12-layer models at 224px input up to 16-layer models with 1024 feature dimensions at 448px input, each trained on either Objaverse alone or Objaverse combined with MVImgNet, with Blender rendering scripts included for data preparation. The code is Apache 2.0 licensed, but the released weights are CC-BY-NC 4.0 and restricted to research use, so commercial products need their own training run. It serves researchers and graphics developers experimenting with fast single-image 3D reconstruction.