NavCrafter: Exploring 3D Scenes from a Single Image

Hongbo Duan; Peiyu Zhuang; Yi Liu; Zhengyang Zhang; Yuxin Zhang; Pengting Luo; Fangming Liu; Xueqian Wang

cs.CVApr 3, 2026

NavCrafter: Exploring 3D Scenes from a Single Image

Authors: Hongbo Duan, Peiyu Zhuang, Yi Liu, Zhengyang Zhang, Yuxin Zhang, Pengting Luo, Fangming Liu, Xueqian Wang

Organizations: Center for Artificial Intelligence and Robotics, Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China · Peng Cheng Laboratory, 518108, China · School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, China · Central Media Technology Institute, Huawei Incorporated Company, China

Abstract

Creating flexible 3D scenes from a single image is vital when direct 3D data acquisition is costly or impractical. We introduce NavCrafter, a novel framework that explores 3D scenes from a single image by synthesizing novel-view video sequences with camera controllability and temporal-spatial consistency. NavCrafter leverages video diffusion models to capture rich 3D priors and adopts a geometry-aware expansion strategy to progressively extend scene coverage. To enable controllable multi-view synthesis, we introduce a multi-stage camera control mechanism that conditions diffusion models with diverse trajectories via dual-branch camera injection and attention modulation. We further propose a collision-aware camera trajectory planner and an enhanced 3D Gaussian Splatting (3DGS) pipeline with depth-aligned supervision, structural regularization and refinement. Extensive experiments demonstrate that NavCrafter achieves state-of-the-art novel-view synthesis under large viewpoint shifts and substantially improves 3D reconstruction fidelity.

Figures & tables

Fig. 2: The NavCrafter framework consists of three modules: (1) Controllable novel-view synthesis via video diffusion, integrating camera trajectories to control video generation and achieve temporally consistent novel views; (2) Iterative view synthesis with collision-aware camera trajectory planning, avoiding scene collisions and optimizing camera trajectories; (3) Geometry-aware 3D reconstruction with enhanced 3D Gaussian Splatting, incorporating depth-aligned supervision, structural regularization and image diffusion model refinement.

Fig. 3: Qualitative comparison with prior methods in controllable novel view synthesis, where the first column shows the input image and camera trajectory. Blue bounding boxes indicate reference areas for easier comparison, while orange ones highlight low-quality generations.

Fig. 4: Qualitative comparison with prior methods in 3D scene reconstruction, where blue bounding boxes show visible regions derived from input image and yellow bounding boxes highlight low-quality regions.

Method	RealEstate10K			DL3DV			Tanks and Temples
Method	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
MotionCtrl [ 8 ]	0.381	16.17	0.443	0.296	15.85	0.485	0.389	14.62	0.421
Scene-Splatter [ 32 ]	0.332	17.91	0.506	0.259	16.80	0.519	0.345	15.20	0.489
ViewCrafter [ 17 ]	0.258	18.52	0.518	0.236	16.98	0.526	0.285	16.38	0.514
Wonderland [ 22 ]	0.206	19.71	0.557	0.218	17.56	0.543	0.221	16.87	0.529
Ours	0.158	21.15	0.680	0.196	17.79	0.576	0.212	17.28	0.547

TABLE I: Quantitative Comparison of Controllable Novel View Synthesis

Method	RealEstate10K		DL3DV		Tanks and Temples
Method	FID $\downarrow$	FVD $\downarrow$	FID $\downarrow$	FVD $\downarrow$	FID $\downarrow$	FVD $\downarrow$
MotionCtrl [ 8 ]	24.12	255.18	28.43	292.62	34.65	327.49
Scene-Splatter [ 32 ]	22.09	223.71	25.70	242.97	26.37	274.25
ViewCrafter [ 17 ]	21.28	208.57	23.46	236.45	24.48	256.13
Wonderland [ 22 ]	16.16	153.48	17.74	169.34	19.46	189.32
Ours	15.88	143.85	16.86	158.61	19.13	181.49

TABLE II: Comparison of Distributional Metrics

Method	RealEstate10K		DL3DV		Tanks and Temples
Method	$R_{err}$ $\downarrow$	$T_{err}$ $\downarrow$	$R_{err}$ $\downarrow$	$T_{err}$ $\downarrow$	$R_{err}$ $\downarrow$	$T_{err}$ $\downarrow$
MotionCtrl [ 8 ]	0.226	0.664	0.343	0.862	0.576	1.207
Scene-Splatter [ 32 ]	0.096	0.280	0.125	0.347	0.241	0.426
ViewCrafter [ 17 ]	0.073	0.194	0.104	0.216	0.144	0.337
Wonderland [ 22 ]	0.046	0.093	0.061	0.130	0.094	0.172
Ours	0.021	0.083	0.047	0.113	0.082	0.148

TABLE III: Comparison of Camera Pose Errors

Method	RealEstate10K			DL3DV			Tanks-and-Temples
Method	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Scene-Splatter [ 32 ]	0.370	16.41	0.482	0.386	15.51	0.503	0.392	15.08	0.479
ViewCrafter [ 17 ]	0.338	16.88	0.523	0.364	15.75	0.529	0.372	15.26	0.491
Wonderland [ 22 ]	0.292	17.15	0.550	0.325	16.64	0.574	0.344	15.90	0.510
Ours	0.179	19.08	0.662	0.291	17.27	0.595	0.308	16.52	0.542

TABLE IV: Quantitative Comparison of 3D Scene Reconstruction

Method	Coverage (%) $\uparrow$	Noise Ratio $\downarrow$	F-score@2cm $\uparrow$	Runtime (min) $\downarrow$
ViewCrafter [ 17 ]	66.20	0.240	0.421	4.75
Ours	77.67	0.078	0.593	4.79

TABLE V: Comparison of reconstruction quality and efficiency between Ours and ViewCrafter.

Fig. 5: Comparison of reconstruction quality between Ours and ViewCrafter.

Fig. 6: Ablation study of 3D scene reconstruction.

Depth-aligned Supervision	Structural Regularization	Refinement	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
$\times$	$\times$	$\times$	0.341	15.42	0.471
$\checkmark$	$\times$	$\times$	0.325	16.12	0.502
$\checkmark$	$\checkmark$	$\times$	0.301	17.14	0.534
$\checkmark$	$\checkmark$	$\checkmark$	0.252	18.45	0.610

TABLE VI: Ablation study results of 3D Scene Reconstruction.

Explore similar work

BetterScene: 3D Scene Synthesis with Representation-Aligned Generative Model
Feb 26, 2026Yuci Han, Charles Toth, John E. Anderson +2Novel View Synthesis Sparse-View 3D Reconstruction
Dept. of Electrical and Computer Engineering, The Ohio State University · USACE ERDC GRL
FreeScale: Scaling 3D Scenes via Certainty-Aware Free-View Generation
Apr 12, 2026Chenhan Jiang, Yu Chen, Qingwen Zhang +4Novel View Synthesis Synthetic Data Augmentation
Hong Kong University of Science and Technology · National Univeristy of Singapore · KTH Royal Institute of Technology +3
GenNVS: Geometry-enhanced Novel View Synthesis via Disentangled 3D Prior
Sep 28, 2026Yajiao Xiong, Youyu Luan, Xiaoyu Zhou +1Video Diffusion Models Novel View Synthesis
Peking University

Feb 26, 2026J/K move · Enter open · S save

BetterScene: 3D Scene Synthesis with Representation-Aligned Generative Model

Yuci Han, Charles Toth, John E. Anderson, William J. Shuart+1

Dept. of Electrical and Computer Engineering, The Ohio State University · USACE ERDC GRL

Novel View Synthesis Sparse-View 3D Reconstruction Video Diffusion Models 3D Reconstruction

Open paper AskPDF arXiv

We present BetterScene, an approach to enhance novel view synthesis (NVS) quality for diverse real-world scenes using extremely sparse, unconstrained photos. BetterScene leverages the production-ready Stable Video Diffusion (SVD) model pretrained on billions of frames as a strong backbone, aiming to mitigate artifacts and recover view-consistent details at inference time. Conventional methods have developed similar diffusion-based solutions to address these challenges of novel view synthesis. Despite significant improvements, these methods typically rely on off-the-shelf pretrained diffusion priors and fine-tune only the UNet module while keeping other components frozen, which still leads to inconsistent details and artifacts even when incorporating geometry-aware regularizations like depth or semantic conditions. To address this, we investigate the latent space of the diffusion model and introduce two components: (1) temporal equivariance regularization and (2) vision foundation model-aligned representation, both applied to the variational autoencoder (VAE) module within the SVD pipeline. BetterScene integrates a feed-forward 3D Gaussian Splatting (3DGS) model to render features as inputs for the SVD enhancer and generate continuous, artifact-free, consistent novel views. We evaluate on the challenging DL3DV-10K dataset and demonstrate superior performance compared to state-of-the-art methods.

Abstract

Explore similar work

BetterScene: 3D Scene Synthesis with Representation-Aligned Generative Model

FreeScale: Scaling 3D Scenes via Certainty-Aware Free-View Generation

GenNVS: Geometry-enhanced Novel View Synthesis via Disentangled 3D Prior