WorldRecon: Generative Multi-View Mesh Reconstruction for Large-Scale 3D World
WorldRecon turns unposed RGB views into coherent large-scene meshes by using a noisy reconstruction scaffold and generative block completion.
The system estimates camera poses and a coarse global point cloud from the input images, then partitions the scene into overlapping blocks. Each block is generated with local point-cloud and image conditioning, producing structure, geometry, and PBR materials. Neighbor-aware latent inpainting and causal world generation carry context across blocks to keep the final mesh unified. The paper reports state-of-the-art results, including lower Chamfer distance and higher F-score than the strongest baselines on synthetic and real indoor scenes. HF Daily Papers' note
The system estimates camera poses and a coarse global point cloud from the input images, then partitions the scene into overlapping blocks. Each block is generated with local point-cloud and image conditioning, producing structure, geometry, and PBR materials. Neighbor-aware latent inpainting and causal world generation carry context across blocks to keep the final mesh unified. The paper reports state-of-the-art results, including lower Chamfer distance and higher F-score than the strongest baselines on synthetic and real indoor scenes. HF Daily Papers' note
score 5