Recent advances in single-image 3D Gaussian head reconstruction have enabled highly realistic and freely renderable digital heads from a single portrait. However, reconstruction and rendering can weaken the forgery traces in the source portrait, making the resulting 3D face difficult to classify whether its underlying face is real or fake, and thereby posing risks to identity authentication and face privacy. To study this problem, we introduce the first large-scale benchmark for this task by collecting real portraits and fake portraits from multiple sources and evaluate representative existing detectors on this benchmark, revealing their lack of explicit mechanisms for retaining fine-grained information and maintaining feature consistency across rendered views. To directly address these two limitations, we propose a detector trained with a two-stage strategy. In Stage I, masked autoencoding encourages the visual backbone to retain the fine-grained appearance information required for local reconstruction, while multi-view contrastive learning enforces feature consistency across rendered views of the same head. Since CLS tokens at different depths exhibit complementary spatial attention patterns, Stage II freezes the adapted backbone and concatenates low-, middle-, and high-level CLS tokens for classification. Experiments show that our method achieves the highest accuracy and ranks first across all reported metrics among the evaluated detectors.
Figures & tables
Figure 1: Task definition. Given a 3D Gaussian head reconstructed from a single portrait image, our task is to determine whether its underlying face is real or fake. At inference time, the authenticity decision is made from a single rendered view of the reconstructed head.
Figure 2: Overview of our 3D Gaussian head dataset.
Figure 3: Overview of our two-stage training framework. In Stage 1, we train the visual backbone with two self-supervised objectives: masked autoencoding models local texture details and subtle Gaussian artifacts, while multi-view contrastive learning suppresses view-specific shortcuts. In Stage 2, low-, middle-, and high-level CLS tokens are normalized and concatenated, then fed into an MLP classification head for real/fake prediction.
Method
2D Split
3D Split
Total
T2I
ExpDriven
FaceSwap
Real
Any3DAvatar
Cap4D
FaceLift
GAGAvatar
LAM
IID
89.32
54.04
71.72
85.05
83.60
77.97
77.20
73.00
80.10
78.60
NPR
88.72
70.04
76.80
84.84
86.60
81.67
82.07
73.57
84.50
81.68
FreqNet
83.48
48.64
54.64
79.04
76.83
71.80
67.07
65.63
71.90
70.65
ForAda
95.76
58.28
79.20
89.15
89.37
79.80
84.13
78.20
85.73
82.49
AIDE
95.12
56.68
68.60
72.56
75.37
71.93
72.17
72.27
73.33
73.01
Table 1: Accuracy comparison across 2D forgery categories and 3D reconstruction methods.
Method
Acc
R.Acc
F.Acc
Macro-F1
AUC
AP(fake)
IID
89.64
88.44
90.85
89.64
95.52
96.68
NPR
80.88
87.29
74.46
80.80
89.24
90.53
FreqNet
80.34
76.81
83.86
80.31
89.31
91.58
ForAda
84.34
76.37
92.31
84.24
94.50
95.70
AIDE
86.63
78.14
95.11
86.53
95.07
94.51
Effort
91.87
87.99
95.74
91.86
98.29
98.53
Table 2: OOD evaluation on the GPT-Image-2 subset.
Method
Acc
R.Acc
F.Acc
Macro-F1
AUC
AP(fake)
IID
78.60
81.15
76.05
78.59
86.65
88.53
NPR
81.68
84.84
78.52
81.66
89.54
91.06
FreqNet
70.65
79.04
62.25
70.44
77.61
79.13
ForAda
82.49
82.44
82.55
82.49
90.50
91.87
AIDE
73.01
72.56
73.47
73.01
81.21
83.23
Effort
85.17
87.32
83.01
85.16
92.98
94.16
Table 3: Comparison with state-of-the-art detection baselines.
Method
Acc
R.Acc
F.Acc
Macro-F1
AUC
AP(fake)
IID
67.64
81.85
53.42
66.97
70.89
75.47
NPR
73.37
83.00
63.74
73.12
79.69
83.06
FreqNet
69.91
71.50
68.33
69.91
75.52
78.18
ForAda
76.59
72.47
80.71
76.55
85.19
86.61
AIDE
59.76
56.77
62.76
59.73
62.22
60.06
Effort
79.43
84.04
74.83
79.39
87.62
89.52
Table 4: OOD evaluation on the HyperSwapper subset.
Method
Acc
R.Acc
F.Acc
Macro-F1
AUC
AP(fake)
IID
59.42
98.15
20.69
52.26
60.93
68.73
NPR
64.65
85.43
43.88
63.06
69.54
72.63
FreqNet
60.42
60.16
60.68
60.42
63.73
61.34
ForAda
55.46
81.47
29.45
52.22
57.15
58.63
AIDE
55.85
51.90
59.79
55.78
56.17
52.76
Effort
60.65
88.98
32.31
57.21
63.83
67.90
Table 5: OOD evaluation on the PersonaLive subset.
Figure 4: Qualitative masked-reconstruction results. With 75% of the input patches masked, the MAE recovers coherent facial structure and fine-grained local details in both real and fake 3D Gaussian-rendered heads.
Method
Acc
R.Acc
F.Acc
Macro-F1
AUC
AP(fake)
IID
82.55
81.91
83.19
82.55
90.16
92.06
NPR
81.83
88.14
75.53
81.76
87.66
90.70
FreqNet
68.87
72.92
64.82
68.82
76.08
77.85
ForAda
83.60
89.85
77.34
83.53
90.78
92.84
AIDE
81.40
89.11
73.69
81.29
87.20
90.19
Effort
85.46
89.57
81.37
85.44
92.16
93.95
Table 6: OOD evaluation on the Any3DAvatar subset.
Strategy
Acc
R.Acc
F.Acc
Macro-F1
AUC
AP(fake)
w/o Stage I
85.79
86.39
85.20
85.79
93.18
94.14
w/o MAE
86.96
88.40
85.52
86.96
93.81
94.69
w/o CL
87.18
88.56
85.80
87.18
94.16
94.85
Ours
89.51
90.84
88.19
89.51
95.36
96.03
Table 7: Ablation study of Stage I training. We report results without masked autoencoding (w/o MAE), without multi-view contrastive learning (w/o CL), and without Stage I training (w/o Stage I). All results are reported as percentages.
Strategy
Intra-Sim ↑
Inter-Sim ↓
Sep-Gap ↑
Real
Fake
Real
Fake
Real
Fake
w/o CL
0.7427
0.7334
0.3499
0.3678
0.3928
0.3656
w/ CL
0.8965
0.8926
0.3293
0.3470
0.5672
0.5456
Table 8: Cross-view representation analysis. Intra-Sim measures similarity across views of the same 3D head, Inter-Sim across different heads of the same authenticity class, and Sep-Gap is their difference.
Figure 5: t-SNE of multi-view representations before (left) and after (right) contrastive learning, showing tighter within-head clusters and better separation across identities.
Figure 6: CLS-token attention across network depths. CLS-token attention maps, averaged over all attention heads, for the selected low-, middle-, and high-level blocks. Different depths exhibit distinct spatial response patterns.
Strategy
Acc
R.Acc
F.Acc
Macro-F1
AUC
AP(fake)
w/ Low-Layer
68.57
69.26
67.87
68.53
75.04
75.72
w/ Mid-Layer
83.16
84.8
81.50
83.15
90.76
92.09
w/ High-Layer
85.64
87.46
83.82
85.64
92.68
93.80
w/ Last-Layer
84.31
85.93
82.69
84.31
91.62
92.95
w/ Multi-Layer
89.51
90.84
88.19
89.51
95.36
96.03
Table 9: Stage II ablation of CLS token selection. We compare the average results of single-layer classifiers in the low-, middle-, and high-depth ranges, the last-layer classifier, and low–middle–high CLS token fusion.
Visual Backbone
Acc
R.Acc
F.Acc
Macro-F1
AUC
AP(fake)
DINOv2 ViT-L/14
86.92
88.20
85.64
86.92
93.35
94.40
DINOv3 ViT-B/16
87.19
90.07
84.31
87.18
94.38
95.14
DINOv3 ViT-L/16
88.49
90.13
86.85
88.49
95.05
95.80
DINOv3 ViT-H+/16
89.51
90.84
88.19
89.51
95.36
96.03
Table 10: Comparison of visual backbone versions and model scales for source-face authenticity detection. DINOv3 ViT-H+/16 achieves the best performance across all metrics.
With the rapid advancement of deep learning in image generation, facial forgery techniques have achieved unprecedented realism, posing serious threats to cybersecurity and information authenticity. Most existing deepfake detection approaches rely on the reconstruction of isolated facial attributes without fully exploiting the complementary nature of multi-modal feature representations. To address these challenges, this paper proposes a novel Multi-Modal 3D Facial Feature Reconstruction Network (M3D-Net) for deepfake detection. Our method leverages an end-to-end dual-stream architecture that reconstructs fine-grained facial geometry and reflectance properties from single-view RGB images via a self-supervised 3D facial reconstruction module. The network further enhances detection performance through a 3D Feature Pre-fusion Module (PFM), which adaptively adjusts multi-scale features, and a Multi-modal Fusion Module (MFM) that effectively integrates RGB and 3D-reconstructed features using attention mechanisms. Extensive experiments on multiple public datasets demonstrate that our approach achieves state-of-the-art performance in terms of detection accuracy and robustness, significantly outperforming existing methods while exhibiting strong generalization across diverse scenarios.
Haotian Wu, Yue Cheng, Shan Bian
a College of Mathematics and Informatics, South China Agricultural University, Wushan 2026 Road, Tianhe District, Guangzhou, 510642, China
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.
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.