Reconstructing high-fidelity 3D head geometry from images is critical for a wide range of applications, yet existing methods face fundamental limitations. Traditional photogrammetry achieves exceptional detail but requires extensive camera arrays (25-200+ views), substantial computation, and manual cleanup in challenging areas like facial hair. Recent alternatives present a fundamental trade-off: foundation models enable efficient single-image reconstruction but lack fine geometric detail, while optimization-based methods achieve higher fidelity but require dense views and expensive computation. We bridge this gap with a hybrid approach that combines the strengths of both paradigms. Our method introduces a multi-view surface normal prediction model that extends monocular foundation models with cross-view attention to produce geometrically consistent normals in a feed-forward pass. We then leverage these predictions as strong geometric priors within an inverse rendering optimization framework to recover high-frequency surface details. Our approach outperforms state-of-the-art single-image and multi-view methods, achieving high-fidelity reconstruction on par with dense-view photogrammetry while reducing camera requirements and computational cost.
Figures & tables
Figure 2: Skullptor reconstructs 3D meshes in two stages. Multi-view normal prediction (Sec. 3.1) produces geometrically consistent surface normals from sparse input images by leveraging cross-view attention across all viewpoints. Mesh optimization (Sec. 3.2) then refines the 3D geometry using the predicted normals as geometric priors within an inverse rendering framework.
Dataset
Method
Avg. Angular Error ↓
Avg. Normal Grad. Err. ↓
<$$$ (%) \uparrow$
<$$$ (%) \uparrow$
<$$$ (%) \uparrow$
Time (s) ↓
Multiface
Sapiens 0.3B
10.3
0.265
61.9
85.3
94.6
7.2
Sapiens 2B
9.23
0.257
69.3
86.3
94.2
41.3
DAViD
9.16
0.250
69.7
86.7
94.7
1.1
Skullptor (Ours)
9.13
0.234
69.2
86.7
94.6
1.5
NPHM
Sapiens 0.3B
7.33
0.194
78.4
93.4
97.8
8.2
Sapiens 2B
6.86
0.185
80.0
94.4
98.3
43.7
Table 1: Normal estimation comparison on Multiface and NPHM datasets.
Figure 3: Qualitative comparison of normal predictions on the NPHM dataset (rows 1–2) and Multiface dataset (rows 3–4).
Dataset
Method
Depth Error
Avg. Angular
Avg. Normal
$<$$$
$<$$$
$<$$$
Runtime
# Views
(mm) ↓
Error ↓
Gradient Error ↓
(%) ↑
(%) ↑
(%) ↑
(min) ↓
Multiface*
Meshroom (Photogrammetry)
0.467
5.43
0.143
85.1
95.8
99.0
7.8
26
2D GS
5.73
10.7
0.206
68.6
85.4
92.5
50
26
SuGaR
5.54
11.5
0.324
66.1
83.8
91.1
42
26
Skullptor (Ours)
2.43
6.17
0.156
91.5
97.6
99.0
0.67
26
Skullptor (Ours, 10 Views)
2.99
6.46
0.157
86.8
95.8
97.9
0.48
10
Table 2: Mesh reconstruction accuracy comparison on Multiface and NPHM datasets.
Figure 4: Qualitative comparison of mesh reconstruction methods on the NPHM (top) and Multiface (bottom) datasets.
Figure 5: Evolution of mesh reconstruction performance with the number of input camera views (NPHM dataset).
Figure 6: Visual comparison of the effect of the number of input views on reconstruction quality for Meshroom and Skullptor.
Normal Metrics
Mesh Metrics
Method
Ang. ↓
Grad. ↓
Depth ↓
Ang. ↓
Grad. ↓
NPHM
DAViD
7.80
0.188
2.68
5.68
0.115
DAViD_ft_mono
8.43
0.174
3.53
6.04
0.116
DAViD_ft_multi
7.13
0.176
2.87
5.98
0.113
Skullptor
7.05
0.167
2.44
5.89
0.110
Multiface
DAViD
8.61
0.238
3.95
6.83
0.163
Table 3: Comparison of normal (left) and mesh (right) error.
Dataset
Normal Estimator
Depth (mm) ↓
Angular Err. ↓
Avg. Normal Grad. Err. ↓
Multiface
Sapiens 0.3B
5.28
8.75
0.177
Sapiens 2B
3.79
6.76
0.167
DAViD
3.74
6.73
0.163
Skullptor (Ours)
3.20
6.45
0.157
NPHM
Sapiens 0.3B
6.38
8.23
0.124
Sapiens 2B
6.76
7.91
0.121
Table S1 : Ablation study on normal estimator impact on final mesh quality (10 views).
Figure S1 : Qualitative impact of normal estimation quality on mesh reconstruction.
Figure S2 : Skullptor results on 12 phone-captured images.
Male
Female
Number of subjects
18
32
Age range
18–64
18–75
Mean age
37.2
42.8
Age < 25
9 (50%)
14 (44%)
Age 25–65
8 (44%)
12 (38%)
Age > 65
1 (6%)
6 (19%)
Table S2 : Subject distribution in the Triplegangers training dataset.
Camera Sampling
Pitch angle θpitch
U(−35∘,35∘)
Yaw angle θyaw
U(−90∘,90∘)
Initial radius r
2.0
Extra 3D translation t
U(−0.2,0.2)3
Scale factor z
U(1.0,1.8)
Projection Matrix
Table S3 : Camera Sampling and Projection Parameters
Figure S3 : Effect of initialization on mesh optimization. Vertex color indicates local edge length (Red short large edges, blue for small edges). Top: Sphere initialization. Bottom: Template mesh initialization. Facial characteristics from the template mesh are discarded, and both initializations converge to comparable results in 300 iterations.
Reconstructing a complete 3D head from a single portrait remains challenging because existing methods still face a sharp quality-speed trade-off: high-fidelity pipelines often rely on multi-stage processing and per-subject optimization, while fast feed-forward models struggle with complete geometry and fine appearance details. To bridge this gap, we propose Any3DAvatar, a fast and high-quality method for single-image 3D Gaussian head avatar generation, whose fastest setting reconstructs a full head in under one second while preserving high-fidelity geometry and texture. First, we build AnyHead, a unified data suite that combines identity diversity, dense multi-view supervision, and realistic accessories, filling the main gaps of existing head data in coverage, full-head geometry, and complex appearance. Second, rather than sampling unstructured noise, we initialize from a Plücker-aware structured 3D Gaussian scaffold and perform one-step conditional denoising, formulating full-head reconstruction into a single forward pass while retaining high fidelity. Third, we introduce auxiliary view-conditioned appearance supervision on the same latent tokens alongside 3D Gaussian reconstruction, improving novel-view texture details at zero extra inference cost. Experiments show that Any3DAvatar outperforms prior single-image full-head reconstruction methods in rendering fidelity while remaining substantially faster.
We propose HeadsUp, a scalable feed-forward method for reconstructing high-quality 3D Gaussian heads from large-scale multi-camera setups. Our method employs an efficient encoder-decoder architecture that compresses input views into a compact latent representation. This latent representation is then decoded into a set of UV-parameterized 3D Gaussians anchored to a neutral head template. This UV representation decouples the number of 3D Gaussians from the number and resolution of input images, enabling training with many high-resolution input views. We train and evaluate our model on an internal dataset with more than 10,000 subjects, which is an order of magnitude larger than existing multi-view human head datasets. HeadsUp achieves state-of-the-art reconstruction quality and generalizes to novel identities without test-time optimization. We extensively analyze the scaling behavior of our model across identities, views, and model capacity, revealing practical insights for quality-compute trade-offs. Finally, we highlight the strength of our latent space by showcasing two downstream applications: generating novel 3D identities and animating the 3D heads with expression blendshapes.
We present a single-image head mesh reconstruction framework that addresses the longstanding challenge of simultaneously preserving facial identity and producing industry-grade topology. Our framework adopts a coarse-to-fine optimization pipeline that refines a rigged template across three stages -- rig, joint, and vertex -- achieving stable convergence and consistent topology. To mitigate the ill-posed nature of single-image 3D face reconstruction and ensure identity preservation, we employ a normal consistency objective jointly with landmark alignment. To further preserve local surface structure and enforce topological regularity, we introduce geometry-aware constraints based on Gaussian curvature and conformal consistency, along with auxiliary regularizations that correct fine artifacts such as lip seams and eyelid discontinuities. Our hierarchical optimization with geometry-aware regularization yields meshes with semantically meaningful edge flow and industry-grade topology. After geometry reconstruction, we extract UV-space texture and normal maps to preserve appearance details for visualization and downstream use. In a user study with 22 professional technical artists, our results were assessed as approaching industry-grade usability, and 95% of participants ranked our method as the top-performing approach, underscoring its effectiveness for real-world digital human production.