Can generative models reconstruct entire 3D scenes from photos?

GenRecon reconstructs detailed 3D scenes from multi-view photos by breaking the scene into overlapping chunks and applying a strong generative model (Trellis.2) to each. A projection-based conditioning mechanism aligns multi-view image features with the generative model, ensuring consistency across views and generating high-fidelity PBR meshes. Results outperform state-of-the-art reconstruction by 16% on indoor scenes.