Extend3D: Town-Scale 3D Generation

Seungwoo Yoon; Jinmo Kim; Jaesik Park

cs.CVMar 31, 2026

Extend3D: Town-Scale 3D Generation

Authors: Seungwoo Yoon, Jinmo Kim, Jaesik Park

Organizations: Seoul National University

Abstract

In this paper, we propose Extend3D, a training-free pipeline for 3D scene generation from a single image, built upon an object-centric 3D generative model. To overcome the limitations of fixed-size latent spaces in object-centric models for representing wide scenes, we extend the latent space in the $x$ and $y$ directions. Then, by dividing the extended latent space into overlapping patches, we apply the object-centric 3D generative model to each patch and couple them at each time step. Since patch-wise 3D generation with image conditioning requires strict spatial alignment between image and latent patches, we initialize the scene using a point cloud prior from a monocular depth estimator and iteratively refine occluded regions through SDEdit. We discovered that treating the incompleteness of 3D structure as noise during 3D refinement enables 3D completion via a concept, which we term under-noising. Furthermore, to address the sub-optimality of object-centric models for sub-scene generation, we optimize the extended latent during denoising, ensuring that the denoising trajectories remain consistent with the sub-scene dynamics. To this end, we introduce 3D-aware optimization objectives for improved geometric structure and texture fidelity. We demonstrate that our method yields better results than prior methods, as evidenced by human preference and quantitative experiments.

Figures & tables

Figure 2: An overall pipeline of our Extend3D. Extend3D consists of two parts: sparse structure generation and structured latent generation. In the denoising part of both steps, an overlapping patch-wise flow was used ( Sec. 4.1 and Fig. 3 ). In sparse structure generation, iterative SDEdit is used to initialize the structure ( Sec. 4.2 ). Vector fields in both steps are optimized with priors ( Sec. 4.3 ).

Figure 3: Overlapping patch-wise flow. The extended latent is divided into latent patches with the sliding window. We then obtain the patch vector for each latent patch and merge them into a single extended latent vector, thereby coupling the patches.

Figure 4: Motivation of under-noising . The blue arrows represent actual noising or denoising, while the purple arrow illustrates how the model is presumed to perceive.

Figure 5: Qualitative result of our Extend3D. Our 3D scene generation result (with $a=b=2$ ) is compared to the results of state-of-the-art 3D generative models. While previous methods may not accurately represent the image or lose scene details, our method effectively expresses the image condition in 3D. The input image is generated using Flux.1 [dev] [ 15 ] . We provide additional results in Sec. A.7 .

Figure 6: Qualitative comparison with SynCity. The results are generated from the text prompt, medieval market .

versus.	Geometry	Faithfulness	Appearance	Completeness
Trellis [ 44 ]	50.0	66.4	67.1	62.1
Hunyuan3D [ 47 ]	73.6	75.7	75.0	75.0
EvoScene [ 48 ]	87.1	87.9	87.1	87.1

Table 1: Human preference win rate (%) of our method.

	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	CD $\downarrow$	F-score (0.05) $\uparrow$
Trellis [ 44 ]	0.650	0.239	10.0	0.0315	0.442
Hunyuan3D [ 47 ]	0.683	0.255	10.4	0.0192	0.567
EvoScene [ 48 ]	0.482	0.310	13.2	0.0188	0.498
Ours w/o $\mathcal{L}_{\mathrm{\textsc{SLat}}}$	0.400	0.333	13.8	0.0078	0.708
Ours	0.240	0.611	20.4	0.0086	0.694

Table 2: Quantitative results.

	CLIP $\uparrow$	HPSv3 $\uparrow$	Intra-LPIPS $\downarrow$
SynCity [ 7 ]	0.251	3.254	0.631
Ours	0.276	3.519	0.571

Table 3: Comparison between 3D scene generation methods.

Figure 7: Ablation study. All the images, except for the ablation of under-noising, are taken from the input image camera viewpoint. We set $a=b=2$ to generate the 3D scenes in this figure.

	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	CD $\downarrow$	F-score (0.05) $\uparrow$
$d=2$	0.251	0.598	19.8	0.0088	0.692
$d=4$	0.240	0.611	20.4	0.0086	0.694
$d=8$	0.237	0.615	20.5	0.0079	0.699

Table 4: Ablation study for varying division factor

d

	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	CD $\downarrow$	F-score (0.05) $\uparrow$
p.w. flow only	0.606	0.209	9.63	0.0348	0.261
+ initialize	0.425	0.312	13.0	0.0083	0.693
+ SS optim.	0.400	0.333	13.8	0.0078	0.708
+ SLat optim.	0.240	0.611	20.4	0.0086	0.694

Table 5: Ablation study on prior initialization and optimization

$t_{\mathrm{noise}}\ /\ t_{\mathrm{start}}$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	CD $\downarrow$	F-score (0.05) $\uparrow$
0.6 / 0.6	0.388	0.324	13.4	0.0081	0.657
0.8 / 0.8	0.550	0.216	9.91	0.0292	0.378
0.8 / 0.6 (over-noise)	0.622	0.219	9.79	0.0518	0.249
0.6 / 0.8 ( under-noise )	0.387	0.327	13.5	0.0078	0.680

Table 6: Ablation study on under-noising.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 13

	VRAM (GB)	Time (min)
Trellis [ 44 ]	17	0.06
Hunyuan3D [ 47 ]	56	1.78
SynCity [ 7 ]	49	52.0
EvoScene [ 48 ]	68	35.3
Ours	28	14.1

Appendix

Table 7: Computational costs of 3D generation methods. (

a=b=2,d=4,n_{\mathrm{iter}}=5

)

	CD $\downarrow$	F-score (0.05) $\uparrow$
$n_{\mathrm{iter}}=0$	0.0079	0.647
$n_{\mathrm{iter}}=1$	0.0075	0.681
$n_{\mathrm{iter}}=2$	0.0077	0.688
$n_{\mathrm{iter}}=3$	0.0082	0.677

Appendix

Table 8: Ablation study on varying number of iterations

n_{\mathrm{iter}}

Figure 8: Qualitative effect of iterative SDEdit

Figure 9: Example results for diverse images. In this figure, we set $a=b=2$ . We compared our results with state-of-the-art open-source 3D generative models. The second image is from CarlaSC dataset [ 41 ] . The other image is generated by ChatGPT [ 30 ] .

Figure 10: Example results for diverse images. In this figure, we set $a=b=2$ . We compared our results with state-of-the-art open-source 3D generative models. The images are generated by ChatGPT [ 30 ] and Flux.1 [dev] [ 15 ] .

Figure 11: Example results for diverse images. In this figure, we set $a=b=2$ . We compared our results with state-of-the-art open-source 3D generative models. The images are generated by ChatGPT [ 30 ] .

Figure 12: Example results for diverse images. In this figure, we set $a=b=2$ . We compared our results with state-of-the-art open-source 3D generative models. The images are generated by ChatGPT [ 30 ] .

Figure 13: Example results for diverse images. In this figure, we set $a=b=2$ . We compared our results with state-of-the-art open-source 3D generative models. The images are generated by ChatGPT [ 30 ] .

Figure 14: Example results for diverse images. In this figure, we set $a=b=2$ . We compared our results with state-of-the-art open-source 3D generative models. The images are generated by ChatGPT [ 30 ] .

Figure 15: Example results of comparison to SynCity. We compared our results with SynCity [ 7 ] . Given a prompt, we first generated an image with ChatGPT [ 30 ] and used Extend3D for our results. The input prompts are on the left.

Figure 16: Example results of comparison to EvoScene [ 48 ] . In this figure, we set $a$ = $b$ = 2. The images are from ChatGPT [ 30 ] and UrbanScene3D [ 19 ] .

Figure 17: The large scale result of Extend3D. We generated large scale ( $a=b=6$ ) 3D scene from the image of Köln captured from Google Earth [ 9 ] . We didn’t use SLat optimization in this result due to the memory shortage.

Figure 18: The large scale result of Extend3D. We generated large scale ( $a=b=6$ ) 3D scene from the image of Athens captured from Google Earth [ 9 ] . We didn’t use SLat optimization in this result due to the memory shortage.

Figure 19: 100 images for appearance evaluation. The images are from ChatGPT [ 30 ] , Flux.1 [dev] [ 15 ] , CarlaSC [ 41 ] , Google Earth [ 9 ] , and UrbanScene3D [ 19 ] .

Figure 20: 45 images for geometry evaluation. The images are from UrbanScene3D [ 19 ] synthetic scenes, together with ground truth meshes.

Image-to-3D Scene Generation

See all 24 papers

Apr 6, 2026

3D-Fixer: Coarse-to-Fine In-place Completion for 3D Scenes from a Single Image

Compositional 3D scene generation from a single view requires the simultaneous recovery of scene layout and 3D assets. Existing approaches mainly fall into two categories: feed-forward generation methods and per-instance generation methods. The former directly predict 3D assets with explicit 6DoF poses through efficient network inference, but they generalize poorly to complex scenes. The latter improve generalization through a divide-and-conquer strategy, but suffer from time-consuming pose optimization. To bridge this gap, we introduce 3D-Fixer, a novel in-place completion paradigm. Specifically, 3D-Fixer extends 3D object generative priors to generate complete 3D assets conditioned on the partially visible point cloud at the original locations, which are cropped from the fragmented geometry obtained from the geometry estimation methods. Unlike prior works that require explicit pose alignment, 3D-Fixer uses fragmented geometry as a spatial anchor to preserve layout fidelity. At its core, we propose a coarse-to-fine generation scheme to resolve boundary ambiguity under occlusion, supported by a dual-branch conditioning network and an Occlusion-Robust Feature Alignment (ORFA) strategy for stable training. Furthermore, to address the data scarcity bottleneck, we present ARSG-110K, the largest scene-level dataset to date, comprising over 110K diverse scenes and 3M annotated images with high-fidelity 3D ground truth. Extensive experiments show that 3D-Fixer achieves state-of-the-art geometric accuracy, which significantly outperforms baselines such as MIDI and Gen3DSR, while maintaining the efficiency of the diffusion process. Code and data will be publicly available at https://zx-yin.github.io/3dfixer.

Ze-Xin Yin, Liu Liu, Xinjie Wang +4

College of Computer Science, Nankai University · Horizon Robotics · D-Robotics +1

3D Scene Completion 3D Asset Generation 3D Diffusion Models

Ask PDF

Jul 6, 2026

SynCity 3000: Bootstrapping Scene-Scale 3D Diffusion

We present SynCity 3000, a framework for generating 3D scenes that are globally coherent while enabling fine-grained layout control. Building on the ability of current image-to-3D generators to produce complex 3D assets from a single image, we extend this capability to the scale of entire scenes by adapting the generator to be applicable as a convolutional operator. We achieve this by fine-tuning the model on scene-like data generated by a new synthetic data engine, which we propose to address the scarcity of 3D scene data for training. The convolutional generator is then applied to a dimetric image of the entire scene, generated from the user prompt, resulting in 3D scenes of arbitrary size and complexity. Across diverse prompts and layouts, SynCity 3000 produces large, coherent, and detailed scenes, addressing the shortcomings of prior approaches to 3D scene generation.

Paul Engstler, Iro Laina, Christian Rupprecht +1

Visual Geometry Group, University of Oxford

3D Scene Generation 3D Scene Layout Generation 3D Diffusion Models Diffusion Models Image-to-3D Scene Generation

Ask PDF

Oct 9, 2026

NOFACE3D: NOise-scheduled Focused Anchoring and Controlled Enhancement for 3D Generation

Recent image-to-3D models generate complete, high-quality assets from a single image, with global shape and appearance that are already convincing; what remains is the fine scale, where small, thin, elongated, or detail-rich regions are reconstructed softly and under-resolved. This gap can arise in generators with fixed-resolution spatial grids: the same spatial sampling density is used for a large, smooth, low-curvature surface and for a thin, high-curvature structure that carries most of the perceived detail. To this end, we introduce NOFACE3D, a training-free framework that refines these detail-critical regions while preserving the rest of the asset. Given a generated mesh, its input image, and a refinement box, NOFACE3D extracts the selected submesh, scales it up to fill the generator's input volume, and re-encodes it into the full-resolution latent grid. Conditioned on the image crop, it refines the region through flow-based sampling with a spatially varying noise schedule that anchors boundary latents. The refined region is then integrated back into the original asset, preserving the surrounding geometry and appearance. We implement NOFACE3D on TRELLIS.2 and its sparse voxel representation; all of it operates purely at inference time on the frozen pretrained generator. We evaluate NOFACE3D on PAniC-3D, TexVerse and BigDetail, a new benchmark of 339 assets with annotated detail-critical regions. On all three, it generates sharper geometry and finer texture detail than the baselines. Project page: https://idkcouscous.github.io/NoFACE3D/

James Rose, Jingwen Liang, Wenchao Ma +2

Roblox · The Pennsylvania State University

3D Asset Generation

Ask PDF

Abstract

Image-to-3D Scene Generation

Explore similar work

3D-Fixer: Coarse-to-Fine In-place Completion for 3D Scenes from a Single Image

SynCity 3000: Bootstrapping Scene-Scale 3D Diffusion

NOFACE3D: NOise-scheduled Focused Anchoring and Controlled Enhancement for 3D Generation