Pretrained 3D generative models produce detailed geometry and appearance but are primarily designed for object-centric generation within a limited spatial extent. Recent approaches address this limitation by partitioning large scenes into smaller spatial regions and applying pretrained 3D generative priors to each region. However, scaling tiled generation to large multi-view scenes makes it challenging to maintain local geometric continuity and global appearance consistency. We present a training-free framework for large-scale textured mesh generation from multi-view images. Our key idea is to scale tiled generation to large scenes with increased spatial detail while coordinating generation both locally and globally. We introduce local context tiled generation to improve geometric continuity between neighboring regions and global appearance alignment to reduce appearance discrepancies across distant regions. An adaptive scene decomposition further determines the number of tiles according to the input scene geometry. Experiments demonstrate improved geometric and appearance fidelity over existing approaches while enabling fine-grained generation of large-scale scenes.
Figures & tables
Figure 2: Overview of our framework. Our method combines geometry-adaptive scene decomposition, local context tiled generation, and global appearance alignment to generate coherent large-scale 3D scene meshes from multi-view images. The tile-wise decoding process shown in the upper right is detailed in Fig. 3 .
Figure 3: Comparison of decoding strategies. Our context-aware tile decoding incorporates neighboring halo features during tile-wise decoding, providing cross-tile context without jointly decoding the entire scene.
Method
# Views
Synthetic
Urban
LPIPS ↓
SSIM ↑
CD ↓
F-score ↑
LPIPS ↓
SSIM ↑
CD ↓
F-score ↑
Hunyuan3D-2.1 [ 8 ]
Single
0.648
0.308
0.016
0.934
0.728
0.162
0.066
0.663
TRELLIS2 [ 41 ]
Single
0.482
0.443
0.023
0.873
0.676
0.206
0.008
0.959
EvoScene [ 51 ]
Single
0.491
0.486
0.023
0.834
0.722
0.174
0.011
0.920
Extend3D [ 45 ]
Single
0.412
0.560
0.024
0.825
0.778
0.250
0.006
0.976
GenRecon [ 30 ]
Multi
0.356
0.635
0.037
0.652
0.747
0.213
0.041
0.695
Table 1: Quantitative comparison with existing 3D generation and reconstruction methods. Best results are shown in bold , and second-best results are underlined .
Figure 4: Qualitative comparison on the NewYork and SanFrancisco scenes. (a) and (b) show zoomed-in views of the upper-right and lower-right regions of each cell, respectively.
Figure 5: Large-scale scene generation with increasing tile resolution ( n=3,5,7 ) over a fixed scene extent. Finer decompositions allocate the pretrained generator’s native resolution to smaller spatial regions, revealing progressively finer geometric and appearance details.
Figure 6: Qualitative ablation of our method. We progressively add (b) cross-tile attention, (c) context-aware decoding, and (d) global alignment to (a) the baseline, improving local continuity and global appearance consistency.
n
# Tiles
LPIPS ↓
# Faces
# SLAT Voxels
3
9
0.683
1,132,659
74,990
5
25
0.620
3,038,380
195,223
7
49
0.605
5,677,351
410,868
Table 2: Scalability analysis with increasing numbers of tiles on the Town scene. The scene extent is fixed across all settings.
Cross-Tile Attention
Context-Aware Decoding
Global Alignment
LPIPS ↓
SSIM ↑
(a)
0.216
0.736
(b)
✓
0.211
0.737
(c)
✓
✓
0.212
0.747
(d)
✓
✓
✓
0.206
0.749
Table 3: Ablation study of the proposed components.
Recent progress in image and video synthesis has inspired their use in advancing 3D scene generation. However, we observe that text-to-image and -video approaches struggle to maintain scene- and object-level consistency beyond a limited environment scale without a persistent, explicit geometric representation. We thus present a geometry-first approach that decouples this complex problem of large-scale 3D scene synthesis into its structural composition, represented as a mesh scaffold, and realistic appearance synthesis, which leverages powerful image synthesis models conditioned on the mesh scaffold. From an input text description, we first construct a mesh capturing the environment's geometry (walls, floors, etc.), and then use image synthesis, segmentation and object reconstruction to populate the mesh structure with objects in realistic layouts. This mesh scaffold is then rendered to condition image synthesis, providing a structural backbone for consistent appearance generation. This enables scalable, arbitrarily-sized 3D scenes of high object richness and diversity, combining robust 3D consistency with photorealistic detail. We believe this marks a significant step toward generating truly environment-scale, immersive 3D worlds.
Recent 3D generative models can synthesize high-quality geometry but often struggle to reproduce intricate textures from reference images, largely due to the scarcity of large-scale 3D training data with rich surface appearance. In contrast, visual generative models are trained on datasets several orders of magnitude larger and excel at modeling complex visual patterns. Motivated by this gap, we introduce Ink3D, a framework that bridges 3D generation with large-scale video generative models to synthesize extremely complex textures. Ink3D first reconstructs a white-mesh geometry using an off-the-shelf 3D generation model. It then employs OrbitPainter, a conditional video generative model, to produce dense orbit-scan videos capturing object appearance across viewpoints. To convert these views into coherent textures, we introduce TextureOptimizer, a neural baking module that integrates dense multi-view observations while mitigating geometry inconsistencies arising from video generation. By decoupling geometry and texture synthesis and leveraging large-scale pretrained video priors, Ink3D enables significantly richer and more faithful texture generation than prior approaches.
Yue Han, Chong Li, Zhening Liu +5
ZGCA & ZGCI · Zhejiang University · Microsoft Research +1
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. Project page: https://kszpxxzmc.github.io/WorldRecon-website/.
Mengchen Zhang, Yang Li, Chunshi Wang +2
Zhejiang University · Tencent Hunyuan · Fudan University