The system uses a VGGT-based encoder, a Perceiver-style compressor to produce a global state, and per-point flow-matching ODE decoding. It adds rendering-based communication guidance so independently flowing points remain part of one coherent surface.
Surflo is useful for feed-forward 3D reconstruction where the number of input views changes and the output resolution should be flexible. The page links to arXiv and code and describes evaluation across several benchmarks plus a new real-world surface dataset.
