zsh523/unirecgen

UniRecGen is a research project focused on turning a small number of photos of an object into a complete 3D model. This is a hard problem because two different approaches each have their own weakness. One approach rebuilds a 3D shape directly from the input photos, which stays faithful to what the camera actually saw but often leaves gaps where the object was not visible. The other approach uses a generative model to imagine the missing geometry, which fills in gaps but can produce a shape that does not quite match across different viewing angles. UniRecGen combines both approaches into one system rather than treating them as separate steps. It first reconstructs the object from the input views and places that partial 3D shape into a shared reference space. It then hands that reconstruction to a generative model, which uses it as a guide to complete and refine the geometry rather than inventing a shape from scratch. According to the project, this combined approach performs better on standard 3D shape benchmarks than several existing methods. The project builds on top of two other open source research systems for extracting geometric information from images and for generating 3D shapes, crediting their authors for laying the groundwork. As of this writing, the authors have published a demo video and a research paper describing the method, but the actual code has not been released yet. The README asks people to star the repository to be notified once the code is published. This repo is therefore useful today mainly as a preview of an upcoming 3D reconstruction and generation tool, not yet as something you can install and run.