Antoine Guédon, Shu Nakamura, Nicolas Dufour, Jiahui Lei, Ko Nishino, Angjoo Kanazawa
Surflo is a feed-forward 3D reconstruction model that compresses a variable number of unposed RGB views into K global latent tokens and independently decodes 3D surface points from noise via flow matching.
Existing feed-forward reconstruction models either output overlapping, unaligned pointmaps per view or are constrained to a fixed, low-resolution output. Optimization-based methods require hundreds of views and are slow.
Surflo uses an encoder to compress variable views into K global latent tokens. The decoder, conditioned on this latent, transports 3D points from noise to the surface via flow matching. Each point is decoded independently, so the number of output points is unconstrained. At inference, a guidance term injects a photometric gradient during ODE integration to enforce consistency among nearby points.
Surflo matches or surpasses feed-forward baselines on surface metrics and is an order of magnitude faster than optimization-based methods that require hundreds of views. It is the first feed-forward approach to combine a global latent with arbitrary-resolution decoding.