Through the use of scattered perspective and color shading techniques, the principal deity in a Thangka is rendered with a sense of three-dimensional depth on a two-dimensional plane, resulting in an artistic characteristic where textural details and geometric structures are highly intertwined. This characteristic leads to a series of issues, such as topological adhesion, loss of detail, blurred textures, geometric distortion, and inconsistent viewpoints, when generic single-view 3D reconstruction approaches are applied directly. To address these issues, this paper proposes a viewpoint-aware and optimized three-plane reconstruction framework called VAWA-3D. During novel view synthesis, the framework strengthens spatial constraints via viewpoint-aware encoding and equips the UNet network with multi-scale adaptive weights to jointly optimize texture details and global geometric structures. In the mesh reconstruction stage, a three-plane consistency verification mechanism is introduced to rectify deviations in geometric mapping. Experiments conducted on the self-built 3D-Thangka dataset and public benchmarks demonstrate that the proposed method outperforms state-of-the-art alternatives by a notable margin on core evaluation metrics with SSIM, LPIPS, and CD values of 0.907, 0.129, and 0.183, respectively, effectively validating the efficiency of our approach in the task of high-fidelity 3D reconstruction of the main figure in Thangka paintings.
近年来,单视图三维重建技术逐步形成了“扩散模型生成多视图[1-5]”与“端到端模型重建网络[6-8]”两阶段主流范式,推动了三维重建从传统几何建模向数据驱动式建模的深刻转变,从图像到三维的跨模态转换已成为文化遗产数字化的主流方向[9-12]。2023年,Liu等[13]提出的Zero-1-to-3(Zero-shot one image to 3D object)首次基于Stable Diffusion框架[14],通过合成数据集微调与CLIP嵌入的双分支编码[15],实现了开放类别的零样本新视图生成,验证了2D扩散模型中的3D几何先验;然而,扩散模型中容易出现视角不一致的问题,为此Shi等[16]提出的Zero123++(a single image to consistent multi-view diffusion base model)进行了专门优化,采用3×2网格拼接、切换线性噪声调度以及缩放参考注意力等策略,优化了多视图联合分布的建模与特征对齐,进而提升了生成结果的连贯性与质量;Wang等[17]提出的CRM则在扩散模型生成多视图的基础上进一步优化,通过引入CCM(Canonical Coordinate Map)补足几何信息,实现了更高质量的纹理细节与几何结构协同重建,丰富了扩散模型生成多视图范式的技术路径。
随着大规模三维数据集[18-21]与Transformer架构[22]的发展,Hong等[23]提出的LRM(Large reconstruction model for single image to 3D)大型重建模型利用Transformer网络将图像tokens映射到隐式三平面,实现了高质量的单视图三维重建;而Tochilkin等[24]提出的TripoSR(Fast 3D object reconstruction from a single image)则在LRM架构的基础上,引入DINOv1训练编码器与多尺度Transformer解码器,通过掩码损失抑制漂浮伪影,最终实现了在0.5s内完成高质量重建,且能直接输出带纹理的网格,较好地平衡了效率与精度。
YuQ, LiX, TangY, et al. Fancy123: One image to high-quality 3D mesh generation via plug-and-play deformation[C]. Proceedings of the Computer Vision and Pattern Recognition Conference. 2025: 595-604.
[2]
LiP, ZhengW, LiuY, et al. Pshuman: Photorealistic single-image 3D human reconstruction using cross-scale multiview diffusion and explicit remeshing[C]. Proceedings of the Computer Vision and Pattern Recognition Conference, 2025: 16008-16018.
[3]
WuK, LiuF, CaiZ, et al. Unique3D: High-quality and efficient 3D mesh generation from a single image[J]. Advances in Neural Information Processing Systems, 2024, 37: 125116-125141.
[4]
WangP, ShiY. Imagedream: Image-prompt multi-view diffusion for 3D generation[PP/OL]. V1. arXiv (2023-12-02) [2026-06-22]. DOI:10.48550/arXiv.2312.02201 .
[5]
TangJ, ChenZ, ChenX, et al. Lgm: Large multi-view gaussian model for high-resolution 3D content creation[C]. European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 1-18.
[6]
WeiX, ZhangK, BiS, et al. Meshlrm: Large reconstruction model for high-quality meshes[PP/OL]. V2. arXiv (2025-01-23) [2026-06-22]. DOI:10.48550/arXiv.2404.12385 .
[7]
WangJ, ChenM, KaraevN, et al. Vggt: Visual geometry grounded transformer[C]. Proceedings of the Computer Vision and Pattern Recognition Conference, 2025: 5294-5306.
[8]
BossM, HuangZ, VasishtaA, et al. Sf3D: Stable fast 3D mesh reconstruction with uv-unwrapping and illumination disentanglement[C]. Proceedings of the Computer Vision and Pattern Recognition Conference, 2025: 16240-16250.
[9]
HuQ, WangJ, PengX, et al. Three-dimensional reconstruction image generation of traditional Chinese painting elements[J]. Engineering Applications of Artificial Intelligence, 2025, 158: 111417.
[10]
ZhangS, WangW, LuB, et al. High-fidelity 3D buddhist sculpture reconstruction from single images using domain-adaptive diffusion[J]. npj Heritage Science, 2025, 13(1): 670.
[11]
DuY, LercariN. Statue Diff: Regionally enhanced cross-scale diffusion for single-view 3D reconstruction of classical statues[J]. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2025, 48: 429-434.
[12]
LiuY, ShenT, HuoL, et al. Dense 3D reconstruction of ancient architectural heritage by integrating cross-view geometric prior and semantic enhancement[J]. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2025, 48: 873-880.
[13]
LiuR, WuR, Van HoorickB, et al. Zero-1-to-3: Zero-shot one image to 3D object[C]. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023: 9298-9309.
[14]
RombachR, BlattmannA, LorenzD, et al. High-resolution image synthesis with latent diffusion models[C]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, 2022: 10684-10695.
[15]
RadfordA, KimJ W, HallacyC, et al. Learning transferable visual models from natural language supervision[C]. International Conference on Machine Learning. PmLR, 2021: 8748-8763.
[16]
ShiR, ChenH, ZhangZ, et al. Zero123++: A single image to consistent multi-view diffusion base model[PP/OL]. V1. arXiv (2023-10-23) [2026-06-22]. DOI:10.48550/arXiv.2310.15110 .
[17]
WangZ, WangY, ChenY, et al. Crm: Single image to 3D textured mesh with convolutional reconstruction model[C]. European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 57-74.
[18]
DeitkeM, SchwenkD, SalvadorJ, et al. Objaverse: A universe of annotated 3D objects[C]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023: 13142-13153.
[19]
WuT, ZhangJ, FuX, et al. Omniobject3d: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation[C]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 803-814..
[20]
DownsL, FrancisA, KoenigN, et al. Google scanned objects: A high-quality dataset of 3D scanned household items[C]. 2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022: 2553-2560.
[21]
ChangA X, FunkhouserT, GuibasL, et al. Shapenet: An information-rich 3d model repository[PP/OL]. V1. arXiv (2015-12-09) [2026-06-22]. DOI:10.48550/arXiv.1512.03012 .
[22]
VaswaniA, ShazeerN, ParmarN, et al. Attention is all you need[PP/OL]. V7. arXiv (2023-08-02) [2026-07-09]. DOI:10.48550/arXiv.1706.03762 .
[23]
HongY, ZhangK, GuJ, et al. Lrm: Large reconstruction model for single image to 3D[C]. International Conference on Learning Representations, 2024: 50678-50702.
[24]
TochilkinD, PankratzD, LiuZ, et al. Triposr: Fast 3D object reconstruction from a single image[PP/OL]. V1. arXiv (2024-03-04) [2026-06-22]. DOI:10.48550/arXiv.2403.02151 .
[25]
RonnebergerO, FischerP, BroxT. U-net: Convolutional networks for biomedical image segmentation[C]. International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer International Publishing, 2015: 234-241.