Any3DAvatar: Fast and High-Quality Full-Head 3D Avatar Reconstruction from Single Portrait Image

Yujie Gao; Yao Xiao; Xiangnan Zhu; Ya Li; Yiyi Zhang; Liqing Zhang; Jianfu Zhang

cs.CVApr 15, 2026

Any3DAvatar: Fast and High-Quality Full-Head 3D Avatar Reconstruction from Single Portrait Image

Authors: Yujie Gao, Yao Xiao, Xiangnan Zhu, Ya Li, Yiyi Zhang, Liqing Zhang, Jianfu Zhang

Organizations: Shanghai Jiao Tong University

Abstract

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.

Figures & tables

Figure 2: Pipeline of Any3DAvatar. Given a single portrait, we extract image and Gaussian tokens, jointly denoise them with a DiT backbone, and decode the outputs into 3D Gaussian point clouds through a few-step feed-forward path at inference. During training, we additionally supervise the same tokens with a view-conditioned image decoding branch.

Figure 3: Overview of the three complementary subsets of our AnyHead dataset.

Figure 4: Qualitative visual comparison between Any3DAvatar and other baselines.

NeRSemble	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	CSIM $\uparrow$	DS $\downarrow$	US $\uparrow$	Ava-256	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	CSIM $\uparrow$	DS $\downarrow$	US $\uparrow$
PanoHead	0.3241	13.0231	0.7446	0.8128	0.1274	4.6874	PanoHead	0.3209	11.3416	0.7641	0.8511	0.1800	5.2024
SphereHead	0.3690	11.6093	0.7265	0.6739	0.2334	1.5416	SphereHead	0.2909	10.9275	0.7793	0.8306	0.1992	4.8452
Arc2Avatar	0.4003	10.4641	0.6567	0.6195	0.2838	2.9166	Arc2Avatar	0.3488	9.2206	0.7176	0.6990	0.3785	2.2292
FaceLift	0.3056	13.4421	0.7558	0.8237	0.1107	3.6458	FaceLift	0.2125	16.1561	0.8123	0.8429	0.0910	6.2024
ID-Sculpt	0.3169	12.0811	0.7425	0.7759	0.1476	1.7916	ID-Sculpt	0.3017	11.0421	0.7682	0.8217	0.1855	5.0000
HQ-Head	0.3661	12.1271	0.7270	0.7636	0.1311	3.1250	HQ-Head	0.2738	13.5515	0.7953	0.8497	0.1807	4.1905

Table 1: Quantitative comparisons on the NeRSemble and Ava-256 datasets.

Figure 5: Qualitative visualization of the effect of different denoising step counts on the multi-view rendering quality of generated 3D heads.

Method	PanoHead	SphereHead	Arc2Avatar
Time (s) $\downarrow$	130.88	127.13	8274.67
FaceLift	ID-Sculpt	HQ-Head	Any3DAvatar
13.15	3198.79	3974.05	0.64

Table 2: Inference time comparison.

Step	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	CSIM $\uparrow$	DS $\downarrow$	Time (s) $\downarrow$
step=1	0.2038	18.6689	0.7541	0.8970	0.0690	0.64
step=5	0.2014	18.6976	0.7545	0.8998	0.0679	3.24
step=10	0.2029	18.6000	0.7537	0.8979	0.0684	6.90
step=20	0.2054	18.5606	0.7529	0.8970	0.0697	13.63
step=30	0.2073	18.6142	0.7529	0.8975	0.0705	19.67

Table 3: Comparison of different denoising steps on the AnyHead dataset.

Panel of Figure 6 — Figure 6: Quantitative comparison of training with and without AVAS on different viewpoint angles (image index).

Figure 7: Visual results of AVAS ablation study. In non-input viewpoints, training with AVAS better preserves texture details in non-input views.

Figure 8: Downstream applications of our method.

Method	NeRSemble
FaceLift	0.3024/13.1803/0.7542/0.8196/0.1116
w/o FaceLift Data	0.2459/16.1976/0.7905/0.8819/0.0659
Method	Ava-256
FaceLift	0.2125/16.1561/0.8123/0.8429/0.0910
w/o FaceLift Data	0.1947/18.3987/0.8370/0.8740/0.0810

Table 4: Comparison with FaceLift using our model trained without FaceLift-generated data. Metrics are reported as LPIPS/PSNR/SSIM/CSIM/DreamSim.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Examples of AI-generated heads used in our dataset construction pipeline.

Figure 10: Examples of digital human heads rendered in UE MetaHuman with diverse appearance attributes and expressions.

Figure 11: Examples of accessory-rich heads generated by controlled multi-view accessory augmentation.

Dataset	ID Number	Accessory	Full Head	Camera View	Stylized Head	Detailed View
Ava-256	256	✗	✓	80	✗	✓
NeRSemble	418	✗	✗	16	✗	✗
RenderMe-360	500	✓	✓	60	✗	✗
AnyHead	1060	✓	✓	80	✓	✓

Appendix

Table 5: Comparison with existing head datasets.

Method	PanoHead	SphereHead	Arc2Avatar	FaceLift	ID-Sculpt	HQ-Head	Any3DAvatar
Render Time (s) $\downarrow$	0.5628	0.7532	$2.0444\times 10^{-3}$	$3.8667\times 10^{-3}$	$3.0602\times 10^{-3}$	$\mathbf{1.8993\times 10^{-3}}$	$4.4328\times 10^{-3}$
Infer + Render Time (s) $\downarrow$	131.4428	127.8832	8274.6720	13.1539	3198.7931	3974.0519	0.6544
File Type	implicit	implicit	gs	gs	mesh	gs	gs

Appendix

Table 6: Comparison of runtime with full-head reconstruction methods.

Method	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	CSIM $\uparrow$	DS $\downarrow$
TRELLIS	0.3254	11.7264	0.6539	0.7643	0.2281
Open-Diffusion-GS	0.3191	12.2962	0.6443	0.7660	0.1680
Any3DAvatar	0.1894	19.8639	0.7632	0.9258	0.0633

Appendix

Table 7: Quantitative comparison with image-to-3D methods on the AnyHead dataset.

Figure 12: Visual comparison with image-to-3D methods.

Strategy	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	CSIM $\uparrow$	DS $\downarrow$
w/o AVAS	0.2014	18.6976	0.7545	0.8998	0.0679
w/ AVAS	0.1894	19.8639	0.7632	0.9258	0.0633

Appendix

Table 8: Ablation study on the effectiveness of the proposed AVAS strategy.

Dataset	PanoHead	SphereHead	Arc2Avatar	FaceLift	ID-Sculpt	HQ-Head	Any3DAvatar
AnyHead	0.5299	0.5696	0.6502	0.8053	0.6799	0.8695	0.9432
NeRSemble	0.6297	0.6324	0.6135	0.6771	0.6090	0.6409	0.9060

Appendix

Table 9: Quantitative comparison using multi-view 2D IoU on the AnyHead and NeRSemble datasets.

Figure 13: Visual comparison with non-full-head 3D head generation methods.

Figure 14: Visual comparison with non-full-head 3D head reconstruction methods.

Method	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	CSIM $\uparrow$	DS $\downarrow$
GPAvatar	0.3233	12.2337	0.7652	0.8175	0.1410
GAGAvatar	0.2968	12.8123	0.7727	0.8327	0.1149
CAP4D	0.3297	11.1998	0.7435	0.8251	0.1479
LAM	0.3070	14.5441	0.7733	0.8098	0.1386
FastAvatar	0.3611	12.5209	0.7328	0.7312	0.1833
Any3DAvatar	0.2357	16.6202	0.7988	0.8905	0.0621

Appendix

Table 10: Quantitative comparison with non-full-head 3D head reconstruction methods on the NeRSemble dataset.

Figure 15: Training convergence of different noise initialization strategies. The variants using pure image-space Gaussian noise or direct latent-space noise fail to converge, whereas incorporating Plücker ray embedding enables stable convergence.

Method	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	CSIM $\uparrow$	DS $\downarrow$
Trained FaceLift	0.2569	14.6061	0.7141	0.8711	0.1509
Any3DAvatar	0.1894	19.8639	0.7632	0.9258	0.0633

Appendix

Table 11: Quantitative comparison with FaceLift retrained and evaluated on the AnyHead dataset.

Figure 16: Visualized accessory rendering results. Our method provides diverse accessory generation capabilities.

Method	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	CSIM $\uparrow$	DS $\downarrow$
Regression (Step=1)	0.2041	18.6431	0.7531	0.8954	0.0693
Endpoint Denoising (Step=1)	0.2038	18.6689	0.7541	0.8970	0.0690
Endpoint Denoising (Step=5)	0.2014	18.6976	0.7545	0.8998	0.0679

Appendix

Table 12: Comparison of direct regression and conditional endpoint denoising under one- and five-step generation on the AnyHead dataset.

Figure 17: We evaluate our method on more stylized heads, including complex lighting, non-frontal viewpoints, anime-style heads, and 3D-stylized heads, demonstrating strong generation quality.

Explore similar work

FFAvatar: Few-Shot, Feed-Forward, and Generalizable Avatar Reconstruction
May 14, 2026Thuan Hoang Nguyen, Jiahao Luo, Yinyu Nie +33D Gaussian Splatting 3D Reconstruction
Snap Inc. · MBZUAI · University of California, Santa Cruz
S-Avatar: Diffusion-Guided Gaussian Head Avatars from a Single Image
Jul 30, 2026Hail Song, Seokhwan Yang, Jiwon Yang +23D Gaussian Splatting 3D Avatar Generation
KAIST UVR Lab · KAIST KI-ITC ARRC
OMG-Avatar: One-shot Multi-LOD Gaussian Head Avatar
Mar 2, 2026Jianqiang Ren, Lin Liu, Steven Hoi3D Avatar Generation 3D Face Reconstruction
Tongyi Lab, Alibaba Group

May 14, 2026J/K move · Enter open · S save

FFAvatar: Few-Shot, Feed-Forward, and Generalizable Avatar Reconstruction

Thuan Hoang Nguyen, Jiahao Luo, Yinyu Nie, Hao Li+2

Snap Inc. · MBZUAI · University of California, Santa Cruz

3D Gaussian Splatting 3D Reconstruction 3D Avatar Generation Sparse-View 3D Reconstruction Human Avatar Animation 3D Face Reconstruction+1

Open paper AskPDF arXiv

Avatar reconstruction has traditionally relied on per-subject optimization that requires hours of computation or on expensive preprocessing that limits scalability. We introduce FFAvatar, a generalizable feed-forward framework that reconstructs high-quality, animatable 3D Gaussian head avatars from few-shot unposed portrait images in seconds. FFAvatar fuses information from multiple source images into a unified canonical Gaussian representation through Multi-View Query-Former, which is animated via FLAME parameters predicted end-to-end directly from pixels, eliminating the overhead of offline FLAME extraction. We further propose a three-stage training curriculum that achieves both broad generalization and high-fidelity reconstruction: (i) scalable pretraining on extensive monocular video data with over 1M identities to learn strong generalizable priors; (ii) multi-view fine-tuning on a small but high-quality dataset of 360-degree captures to enhance geometric fidelity and extreme-view awareness; and (iii) optional personalization that adapts to specific identities for maximum fidelity within 500 optimization steps. Extensive experiments demonstrate that FFAvatar sets a new standard for identity preservation, geometric consistency, and animation fidelity. On the NeRSemble benchmark, it outperforms the state-of-the-art LAM by a substantial 5.5 PSNR gain. Furthermore, FFAvatar enables real-time deployment, reconstructing avatars in 2 seconds without personalization and 10 seconds with personalization, while supporting 49 FPS animation on a single NVIDIA A100 GPU.

Abstract

Explore similar work

FFAvatar: Few-Shot, Feed-Forward, and Generalizable Avatar Reconstruction

S-Avatar: Diffusion-Guided Gaussian Head Avatars from a Single Image

OMG-Avatar: One-shot Multi-LOD Gaussian Head Avatar