SparseFusion: Distilling View-Conditioned Diffusion for 3D Reconstruction

Figure 1 . Sparse-view Reconstruction. We present SparseFusion, an approach for 3D reconstruction given a few (e.g. just two) segmented input images with known relative pose. SparseFusion is able to generate a 3D consistent neural scene representation, enabling us to render novel views and extract the underlying geometry, while being able to generate detailed and plausible structures in uncertain or unobserved regions (e.g. front of hydrant, teddy's face, back of laptop, or left side of toybus). Please see project page for 360-degree visualizations.