Geometry-Aware Representation Denoising for Robust Multi-view 3D Reconstruction

Researchers introduced GARD, a diffusion-based framework for robust multi-view 3D reconstruction.

It denoises feature representations inside a feed-forward reconstructor to better restore scene geometry.

An RGB decoder is added to recover sharper images alongside 3D structure.

On Depth Anything 3 benchmark tests, the method performed strongly on degraded real-world inputs.