A new arXiv paper introduces WorldRecon, a generative multi-view reconstruction framework that recovers complete, large-scale 3D scene meshes from unposed RGB images. It uses a coarse, noisy global point cloud and camera poses estimated from the inputs as a global scaffold, partitions the scene into overlapping spatial blocks, and equips the native 3D generative model TRELLIS.2 with reconstruction-based conditioning to produce sparse structure, detailed geometry and PBR materials from each block's local point-cloud crop and relevant image observations. To keep scenes coherent, the authors add neighbour-aware latent inpainting and a causal world-generation procedure that propagates structural, geometric and material context across blocks. The paper reports state-of-the-art results, reducing Chamfer distance by 14.9% and 43.1% and improving F-score by 17.3% and 12.0% over the strongest baselines on synthetic and real-world indoor scenes respectively, with strong claimed sim-to-real generalization.

WorldRecon·TRELLIS.2·arXiv