Developers can create novel-view renderings from posed or unposed images and investigate large-scale feed-forward reconstruction. Core capabilities include Multiple input images, Optional camera poses, 3D Gaussian output, and Frozen geometry branch.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Learned Gaussian attributes, Opacity-based artifact suppression, Large-viewpoint rendering, and Public code and live demo.
