4 input views
Loading the 3D scene…
Dashed GT cameras are aligned once to the recovered scene. Reprojection uses the predicted cameras.
Generative Structure from Motion with Native 3D Diffusion
Geometry, appearance, and cameras. Recovered together.Explore in 3D
INTERACTIVE RESULTS
Explore in 3D.
Rotate the scene. Pick an input view.
See how the reconstruction lines up.
From four views to 3D.
How it works
Quantitative results
Full results All metrics
Citation