Creating high-fidelity, animatable 3D avatars from a single image remains a formidable challenge. We identified three desirable attributes of avatar generation: 1) the method should be feed-forward, 2) model a 360° full-head, and 3) should be animation-ready. However, current work addresses only two of the three points simultaneously. To address these limitations, we propose OMEGA-Avatar, the first feed-forward framework that simultaneously generates a generalizable, 360°-complete, and animatable 3D Gaussian head from a single image. Starting from a feed-forward and animatable framework, we address the 360° full-head avatar generation problem with two novel components. First, to overcome poor hair modeling in full-head avatar generation, we introduce a semantic-aware mesh deformation module that integrates multi-view normals to optimize a FLAME head with hair while preserving its topology structure. Second, to enable effective feed-forward decoding of full-head features, we propose a multi-view feature splatting module that constructs a shared canonical UV representation from features across multiple views through differentiable bilinear splatting, hierarchical UV mapping, and visibility-aware fusion. This approach preserves both global structural coherence and local high-frequency details across all viewpoints, ensuring 360° consistency without per-instance optimization. Extensive experiments demonstrate that OMEGA-Avatar achieves state-of-the-art performance, significantly outperforming existing baselines in 360° full-head completeness while robustly preserving identity across different viewpoints.
Figures & tables
Figure 1. Given a single portrait image across diverse styles (left), our method performs one-shot reconstruction of a high-fidelity full-head Gaussian avatar (middle). The reconstructed avatars include the back of the head and are capable of high-quality animations (right). teaser
Method
Representation
One-Shot
360∘ Full-Head
Animatable
Rodin ( Wang et al., 2023 )
Tri-plane
✓
✓
✗
PanoHead ( An et al., 2023 )
Tri-plane
✗ ∗
✓
✗
SphereHead ( Li et al., 2024 )
Tri-plane
✗ ∗
✓
✗
FaceLift ( Han et al., 2024 )
3DGS
✓
✓
✗
FATE ( Zhang et al., 2025 )
3DGS
✗
✓
✓
GAGAvatar ( Chu and Harada, 2024 )
3DGS
✓
✗
✓
Table 1. Comparison of properties with state-of-the-art methods. Our method is the first framework for one-shot reconstruction of animatable 360∘ full-head avatars with 3D Gaussian Splatting.
Figure 2. Pipeline Overview. Given the source and target images, we leverage diffusion models to synthesize multi-view RGB images and corresponding normal maps. These normal maps are used for semantic-aware mesh deformation, while pixel-wise features are extracted from multi-view RGB images. Multi-view features are subsequently aggregated into a canonical UV feature map through the multi-view feature splatting module. The UV features and vertex features extracted from the deformed mesh are decoded and anchored to the mesh via UV mapping. For animation, the expression and pose derived from the target image are injected into the deformed mesh. Finally, the rendered output is enhanced by a neural refiner to generate the final full-head avatar. A schematic pipeline diagram illustrating the full-head avatar generation process. The diagram shows source and target images as inputs, followed by diffusion-based synthesis of multi-view RGB images and normal maps. The normal maps guide mesh deformation, while features extracted from the multi-view images are aggregated into a canonical UV feature map through a feature splatting module. The UV features and mesh vertex features are decoded and mapped onto the deformed mesh. Expression and pose information from the target image are then applied to the mesh for animation, and the final rendered result is refined by a neural refinement module.
Figure 3. Semantic-aware Mesh Deformation. Direct optimization with normal guidance disrupts FLAME’s parametric structure, causing severe surface irregularities and topological artifacts (left). Note the cranial holes and degenerated facial features like eyes and ears (highlighted in red boxes). Our approach (right) mitigates this via semantic-aware topology preservation and a semantic-aware Laplacian. We preserve FLAME’s clean topology and ensure 360∘ geometric consistency, enabling fine detail generation without compromising facial structural integrity.
Figure 4. Multi-view Feature Splatting. Taking a frontal view as an example, we first obtain per-pixel UV coordinates and normals via rasterization, mapping features to UV space using differentiable bilinear splatting. Our hierarchical UV mapping builds a multi-resolution pyramid to fill missing regions in a coarse-to-fine manner. Simultaneously, we calculate a fusion weight map by combining view-dependent confidence and UV sampling density. Finally, the visibility-aware fusion module aggregates weighted features across all views to generate the final UV feature map. Diagram illustrating the per-view feature processing pipeline. For each input view, rasterization yields per-pixel UV coordinates, followed by bilinear splatting of 2D features onto hierarchical UV maps. The resulting partial UV feature maps are merged via visibility-aware fusion to form a complete canonical UV feature map.
Figure 5. Novel view synthesis from single image on the Avatar-256 dataset. Compared to state-of-the-art methods, our approach produces high-quality rendering results, even under unseen and extreme side-view facial angles. Note that PanoHead and SphereHead require inputs aligned to the FFHQ canonical space, leading to apparent scale differences. We use the red boxes to highlight the visual artifacts. A qualitative comparison showing rendered head avatars from multiple methods under extreme side-view facial angles. The visual results illustrate differences in identity consistency and rendering quality across methods, with the proposed approach producing more stable facial structure and fewer visual artifacts in challenging viewpoints.
Avatar-256 Novel Views
NeRSemble Novel Views
Method
PSNR ↑
SSIM ↑
LPIPS ↓
CSIM ↑
DS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
CSIM ↑
DS ↓
PanoHead
17.995
0.5932
0.2592
0.4556
0.1663
17.151
0.6098
0.2706
0.5742
0.1082
SphereHead
18.932
0.6238
0.2436
0.4714
0.1509
17.446
0.6183
0.2628
0.5781
0.1053
GAGAvatar
22.802
0.7714
0.1719
0.4682
0.1762
22.556
0.8027
0.1505
0.6867
0.1033
SOAP
21.335
0.7287
0.1948
0.5783
0.1890
19.539
0.7471
0.1975
0.6855
0.1128
LAM
21.802
0.7314
0.1997
0.4842
0.1818
20.263
0.7465
0.1853
0.6321
0.1118
Table 2. Quantitative Comparison of Novel View Synthesis on the Avatar-256 and NeRSemble Datasets. Colors denote the best and second-best
Figure 6. Qualitative Ablation on the NeRSemble dataset. We compare results for models trained: (1) without multi-view diffusion, (2) without using the semantic-aware Laplacian term, (3) without differentiable bilinear splatting, (4) without hierarchical UV mapping, (5) without visibility-aware fusion. Ablation
Method
PSNR ↑
SSIM ↑
LPIPS ↓
CSIM ↑
AED ↓
APD ↓
AKD ↓
SOAP
14.557
0.599
0.305
0.662
0.175
0.0562
4.997
Ours (w/o refiner)
18.062
0.714
0.188
0.664
0.116
0.0207
2.196
Ours (full)
20.739
0.744
0.169
0.660
0.109
0.0188
2.023
Table 3. Reenactment evaluation on the VFHQ dataset.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
CSIM ↑
AED ↓
APD ↓
AKD ↓
SOAP
14.607
0.711
0.283
0.735
0.166
0.0351
3.579
Ours (w/o refiner)
19.854
0.765
0.166
0.757
0.107
0.0090
1.551
Ours (full)
22.594
0.820
0.123
0.752
0.097
0.0091
1.515
Table 4. Reenactment evaluation on the HDTF dataset.
Ablation Novel Views
Method
PSNR ↑
SSIM ↑
LPIPS ↓
CSIM ↑
DS ↓
w/o MV Diffusion
22.687
0.774
0.151
0.655
0.124
w/o S.a Laplacian
22.988
0.782
0.150
0.667
0.105
w/o D.B Splatting
23.187
0.789
0.141
0.643
0.098
w/o Hierarchical UV
22.997
0.785
0.149
0.655
0.103
w/o V.a fusion
23.121
0.793
0.151
0.597
0.110
Table 5. Ablation of different components on the NeRSemble dataset.
Figure 7. Additional Results on In-the-wild Images. Our method demonstrates great generalization ability and robustness towards in-the-wild images, and provides realistic unseen view rendering results. Red boxes highlight visual artifacts.
Figure 8. More Visual results on the NeRSemble dataset. Taking a single view as input, we perform 360∘ novel view synthesis to compare our method with state-of-the-art approaches. The results show that our method achieves superior multi-view consistency and accurately reconstructs unseen regions. We use the red boxes to highlight the visual artifacts.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Method
PSNR ↑
SSIM ↑
LPIPS ↓
CSIM ↑
AED ↓
APD ↓
AKD ↓
GAGAvatar ( Chu and Harada, 2024 )
21.83
0.818
0.122
0.816
0.111
0.135
3.349
LAM ( He et al., 2025 )
22.65
0.829
0.109
0.822
0.102
0.134
2.059
SOAP ( Liao et al., 2025 )
14.557
0.599
0.305
0.662
0.175
0.0562
4.997
Ours (w/o refiner)
18.062
0.714
0.188
0.664
0.116
0.0207
2.196
Ours (full)
20.739
0.744
0.169
0.660
0.109
0.0188
2.023
Appendix
Table 6. Self-reenactment comparison on the VFHQ dataset. Colors denote the best and second-best results.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
CSIM ↑
AED ↓
APD ↓
AKD ↓
GAGAvatar ( Chu and Harada, 2024 )
23.13
0.863
0.103
0.862
0.110
0.111
2.985
LAM ( He et al., 2025 )
23.43
0.873
0.097
0.865
0.101
0.093
1.965
SOAP ( Liao et al., 2025 )
14.607
0.711
0.283
0.735
0.166
0.0351
3.579
Ours (w/o refiner)
19.854
0.765
0.166
0.757
0.107
0.0090
1.551
Ours (full)
22.594
0.820
0.123
0.752
0.097
0.0091
1.515
Appendix
Table 7. Self-reenactment comparison on the HDTF dataset. Colors denote the best and second-best results.
Index
NeRSemble ID
Avatar-256 ID
1
017
20220614--1135--DNM410
2
070
20220808--0809--DPE040
3
124
20230308--1352--BDF920
4
175
20230328--0800--BLY735
5
214
20230405--1635--AAN112
6
218
20230726--1657--AYE877
Appendix
Table 8. Identities used from the NeRSemble and Avatar-256 datasets.
Figure 9. Qualitative stress test for extreme-expression reenactment. The top row shows six driving images spanning joy, anger, rage, a manic grin, shock, and surprise, while the leftmost column shows the source portraits used to reconstruct the avatars. A grid with six driving expressions in the top row and ten source portraits in the left column. Each interior cell shows the corresponding source avatar reenacted with the expression of the driving image.
Figure 10. Additional qualitative results on the NeRSemble and Avatar-256 datasets. We use the red boxes to highlight the visual artifacts. Additional qualitative results on the NeRSemble and Avatar-256 datasets
Figure 11. Full-head reconstruction results on in-the-wild images Full-head reconstruction results on in-the-wild images
Building one-shot 3D animatable head avatars is an important yet challenging problem. Existing methods generally collapse under large camera pose variations, compromising the realism of 3D avatars. In this work, we propose a new framework to tackle the novel setting of one-shot 3D full-head animatable avatar reconstruction in a single forward pass via inpainted UV-space Gaussian modeling, enabling 360∘ rendering views and real-time animation. To facilitate efficient animation control, we model 3D head avatars with Gaussian primitives embedded on the surface of a parametric face model within the UV space, and project the input image features to the UV space, resulting in incomplete local UV feature maps. To inpaint the missing regions, we obtain knowledge of full-head geometry and textures from rich 3D full-head priors within a pretrained 3D generative adversarial network (GAN) for global full-head feature extraction and multi-view supervision. Specifically, to enhance the fidelity of 3D reconstruction during inpainting, we take advantage of the symmetric nature of the UV space and human faces to fuse incomplete yet detailed local UV feature maps with the extracted global full-head textures, resulting in inpainted UV Gaussian attribute maps for avatar modeling. Extensive experiments demonstrate that our method is the first to achieve high-quality 3D full-head animatable avatar modeling, significantly improving side and back views while outperforming state-of-the-art animation approaches, thereby improving the realism of 3D animatable avatars.
Shuling Zhao, Dan Xu
The Hong Kong University of Science and Technology · Zeekr Automobile R&D Co., Ltd.
We propose S-Avatar, a novel method for generating photorealistic 3D head avatars from a single image using a diffusion-guided 3D model generation module and strategies for animating 3D Gaussian Splatting (3DGS). While single-image head avatar reconstruction is crucial for lifelike Virtual Reality (VR) applications, existing approaches often struggle to preserve 3D consistency under unseen viewpoints. S-Avatar addresses this limitation through a three-stage pipeline. First, a high-resolution 3DGS is synthesized directly from a single image using a diffusion-based Gaussian splat generation module. Next, the parametric head model FLAME is aligned with the generated 3DGS by optimizing its parameters and spatial transformations. Finally, to adapt the 3DGS to FLAME variations, we construct a binding template that encodes the spatial relationship between the initial splats and FLAME. The dynamic 3D head avatar can then be rendered in real time by deforming the 3DGS with the binding template. By combining diffusion-guided canonical 3DGS generation with FLAME-based control, our method achieves efficient and accurate reconstruction with enhanced 3D consistency. Evaluations on public datasets demonstrate that S-Avatar outperforms state-of-the-art methods in novel-view and expression generation, achieving superior realism and consistency. Consequently, our approach represents a significant advance in accessible avatar creation, applicable to a wide range of VR/AR applications. The project page is available at https://github.com/hailsong/savatar.
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.