Gen3R: Researchers Unify 3D Scene Generation and Feed-Forward Reconstruction in a Single Model

Researchers from Zhejiang University and ByteDance have introduced Gen3R, a method that bridges foundational reconstruction models and video diffusion models to generate 3D scenes from single or multiple images. The approach works by adapting the VGGT reconstruction model to produce geometric latents that are aligned with the appearance latents of pre-trained video diffusion models, enabling joint generation of RGB video and 3D geometry — including camera poses, depth maps, and global point clouds — within a unified latent space.

Accepted to CVPR 2026, Gen3R claims state-of-the-art results in both single- and multi-image conditioned 3D scene generation. The tight coupling of reconstruction and generative priors also improves reconstruction robustness, suggesting mutual benefits for both tasks. Code and an arXiv preprint are available.