Surflo v0
Surflo reconstructs a static scene from multiple RGB images without supplied camera poses. It compresses the views into one representation, then generates oriented surface points that can be converted into a mesh.
Research release
One scene representation supports different output sizes
The model stores 128 latent tokens, regardless of the number of input views. It can decode different numbers of surface points from that same representation. The published input scope is 2–80 views of one static scene.Surflo v0: model architecture and inherited license terms
Portrait images change the reconstruction frame
The implementation rotates portrait images into landscape orientation. If the first image is rotated, the output mesh or point cloud uses that rotated camera frame. This matters when aligning the result with another scene.Surflo: research contributors and implementation
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | AntoineGuedon/Surflo-v0 |
| Contributors | LIX at École polytechnique, Kyoto University, Kyutai and UC Berkeley.Surflo: research contributors and implementation |
| Parameters | 1.39B total; 128.8M trained parameters and a frozen VGGT 1B backbone.Surflo v0: model architecture and inherited license terms |
| Terms | CC BY-NC 4.0 weights; non-commercial Gaussian-Splatting terms for code.Surflo v0: model architecture and inherited license terms |
Related models
A single-image model that generates new scene views from a target camera pose at 256-by-256 resolution.
Launch date not confirmedResearch releaseDownloadable