GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction

Researchers introduced GenRecon, a 3D scene reconstruction method that uses generative priors to better recover large indoor scenes from RGB images.

It splits scenes into overlapping chunks and applies projection-based conditioning to align multi-view features with the 3D generative model.

The approach improves scene-scale reconstruction quality and completeness, with a reported 16% gain over leading methods.

GenRecon also produces editable PBR meshes, making the results more practical for indoor environment modeling.