WorldRecon | Generative Multi-View Mesh Reconstruction

Abstract

Reconstructing complete, high-quality meshes of large-scale scenes from unposed multi-view images requires faithful geometric grounding while plausibly completing surfaces that are unobserved or only sparsely captured. We present WorldRecon, a generative multi-view reconstruction framework for recovering complete, large-scale 3D scene meshes from unposed RGB images. WorldRecon uses a coarse, noisy global point cloud and camera poses estimated from the input images by a reconstruction model as a global scaffold for whole-scene generation and partitions it into overlapping spatial blocks. To robustly leverage these imperfect geometric priors, we equip the native 3D generative model TRELLIS.2 with a dedicated reconstruction-based conditioning architecture that generates each block from its local point-cloud crop and block-relevant image observations, producing sparse structure, detailed geometry, and PBR materials. To maintain scene-level coherence, we introduce neighbour-aware latent inpainting and a causal world-generation procedure that propagates structural, geometric, and material context across overlapping blocks. Together, these designs scale multi-view generative reconstruction to large-scale scenes, yielding coherent, unified meshes. Extensive experiments demonstrate that WorldRecon achieves state-of-the-art reconstruction performance, reducing Chamfer distance by 14.9% and 43.1% while improving F-score by 17.3% and 12.0% over the strongest baselines on synthetic and real-world indoor scenes, respectively, thereby demonstrating faithful and complete generative reconstruction as well as strong sim-to-real generalization.

Method Overview

WorldRecon reconstructs complete scene meshes from unposed multi-view RGB images.

Single-block generation architecture: global point-cloud reconstruction and slicing prepare block-local inputs; image and camera tokens condition three cascaded flow models, with additional point-cloud conditioning for sparse structure. — **Reconstruction-Based Generation.** For each block, DINOv3 image tokens and Plucker-ray camera tokens are fused into fixed-length multi-view tokens, while a point-cloud encoder extracts geometric tokens. Three cascaded reconstruction-based flow models progressively generate the sparse structure, detailed geometry, and PBR materials, which are robust to noisy point clouds and inaccurate camera poses.

World Generation Pipeline

World Generation Pipeline. Guided by a global point-cloud scaffold, WorldRecon completes sparse structure across overlapping blocks, then generates geometry and PBR materials in the same causal order. Neighbour-aware latent inpainting and overlap blending produce a coherent scene mesh.

00:43 · 1440 × 1080 · 30 fps Download video

Qualitative Comparison

More complete surfaces and consistent geometry from the same visual observations.

1 / 5

Quantitative Results

Evaluation on 30 synthetic scenes from 3D-FRONT and 30 real-world scenes from ScanNet++.

3D-FRONT: evaluation on 30 synthetic indoor scenes
Method	3D Geometry			Coverage	Depth				Normal
Method	CD ↓	F-score ↑	NC ↑	Complete. ↑	RMSE ↓	AbsRel ↓	LPIPS ↓	CLIP ↑	Angular ↓	LPIPS ↓	CLIP ↑

30 synthetic indoor scenes from 3D-FRONT.Best in bold; second-best underlined.

Full evaluation and additional metrics in the paper

Large-Scale Scene Reconstruction

Causal world generation extends reconstruction to coherent, multi-room environments.

Reconstructed multi-room apartment alongside detailed PBR and geometry views of the living room and bedroom.

Comparison with Agentic Baseline

Given the same 32 views, WorldRecon preserves scene layout and object detail more faithfully than the agent baseline.

Two indoor scenes compared across Ground Truth, WorldRecon, and GPT-6 Astra Agent. WorldRecon reconstructions retain more scene structure and object detail than the agent outputs.

Scene Reconstruction Results

Complete scene meshes from synthetic images, real-world captures, and generated videos.

Input Views

32 views

Reconstructed Mesh

Interior Views

Surface normals of reconstructed furniture and walls — Surface normals

Depth map of the reconstructed indoor scene — Depth

Reconstructed PBR materials under changed lighting — Relighting

View 1 / 4

Citation

@misc{zhang2026worldrecongenerativemultiviewmesh,
  title         = {WorldRecon: Generative Multi-View Mesh Reconstruction for Large-Scale 3D World},
  author        = {Mengchen Zhang and Yang Li and Chunshi Wang and Tong Wu and Chunchao Guo},
  year          = {2026},
  eprint        = {2610.18846},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.18846}
}