Flow Matching for Efficient Arbitrary-Resolution 3D Surface Reconstruction from Multi-View Images
Surflo addresses a fundamental inefficiency in feed-forward 3D reconstruction by compressing multiple unposed RGB views into a compact global latent representation, then decoding arbitrary-resolution point clouds through flow matching. Unlike existing methods that either produce linearly-growing per-view outputs or commit to fixed resolutions, Surflo generates 1,000 to 1,000,000 surface points from the same latent in a single forward pass while maintaining consistency through inference-time photometric guidance. The approach achieves competitive surface accuracy with an order-of-magnitude speedup over optimization-based methods requiring hundreds of input views.