GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction

rsasaki0109 · x · 2026-08-24

GenRecon introduces a novel approach for high-fidelity 3D reconstruction from multi-view RGB images by tightly coupling it with strong 3D generative priors. It casts reconstruction as conditional generation over spatially-localized, overlapping chunks. By using a projection-based conditioning mechanism aligned with models like Trellis.2, it achieves high fidelity and completeness independent of view ordering.

Original post →

More from Research

Research channel →