Reconstructing animatable 3D human avatars from images is important for applications such as gaming and augmented reality. Feedforward methods make avatar reconstruction accessible by reducing inference time to a few seconds, but recent approaches rely heavily on scaling model capacity, training data, and computation. We ask whether such scaling is necessary. To this end, we introduce ARG-Avatar, an efficient feedforward framework with only 68M trainable parameters that achieves SOTA reconstruction quality. Our key insight is that 3D animation and rendering already provide strong geometric correspondences that existing methods largely attempt to learn from data. In light of this, we design the Foreground Avatar Coordinate RoPE attention (FACRoPE) and Intermediate Token Rendering (ITR) pipeline to guide avatar tokens to attend to their corresponding patches. Experiments across multiple datasets show that ARG-Avatar consistently outperforms prior methods when using two or more input images. Compared with LHM++, ARG-Avatar uses approximately 9× fewer trainable parameters, 4.8× fewer training sequences, and 14× less training compute, while running 6× faster and achieving a 2.7 dB PSNR improvement on the out-of-distribution NeuMan benchmark.
Figures & tables
Figure 2: Out-of-distribution NeuMan PSNR vs. training compute. Circle area is proportional to trainable parameters.
Figure 3: Comparison of token-to-image attention using FACRoPE ( Sec. 3.3 ) and naive attention. From left to right: input images, the reconstructed canonical UV 3DGS avatar, the marked token’s attention over image features, and FAC visualizations in three views.
Figure 4: ARG-Avatar network pipeline. Each input image is processed by SAM3DBody and a ConvNeXt encoder to obtain its pose and features. ARG-Avatar then applies two MV-ARGBlocks to attend to and fuse features from multiple input images. The output tokens are decoded into a canonical Gaussian avatar and pose-dependent deformations for each driving pose.
Figure 5: Pipeline of the Animation-Rendering Guidance Block (ARGBlock) with Intermediate Token Rendering (ITR). UV tokens are decoded into candidate Gaussian avatars, which are then animated and rendered to obtain FAC that guides attention to image features.
NeuMan
Multiview Videos
3D Scans
HumanVid
Method
V
Novel Poses
Input Poses
Novel Poses
Novel Views
Novel Poses
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
SIGMAN [ 102 ]
1
19.62
0.957
0.066
19.96
0.959
0.063
16.52
0.944
0.076
15.27
0.908
0.112
16.17
0.862
0.162
IDOL [ 118 ]
1
24.96
0.964
0.036
25.35
0.966
0.033
23.18
0.962
0.051
21.32
0.935
0.086
20.53
0.899
0.129
LHM [ 67 ]
1
26.10
0.968
0.028
28.40
0.972
0.019
25.30
0.965
0.041
22.81
0.940
0.078
21.23
0.906
0.111
LHM++ [ 68 ]
1
27.71
0.970
0.023
30.62
0.974
0.017
25.34
0.966
0.039
22.31
0.939
0.078
22.03
0.910
0.104
Table 1: Comparison of ARG-Avatar with the baselines. We report novel-pose synthesis on video datasets, novel-view synthesis on 3D scans, and input-pose reconstruction on NeuMan. “Multiview Videos” combines DNARendering and 4D-Dress; “3D Scans” averages results across 2K2K, THuman2.1, and Human4DiT. NeuMan is unseen during training. Other datasets use validation splits constructed according to Sec. 4.1 . First , second , and third indicate the top three results in each metric column.
Figure 6: Qualitative comparison of ARG-Avatar with the baselines. All samples are unseen during ARG-Avatar training, but 2K2K is used to train LHM and LHM++. LHM uses the first view as input; LHM++ and ARG-Avatar use all six views. The first column of each group shows an overlay of the rendered and target images. Orange and blue boxes highlight differences.
Figure 7: Ablation study of ARG-Avatar’s three key components: FACRoPE attention, the ITR pipeline, and pose-dependent deformation (“Deform”). “L3” denotes a variant with one additional layer. The star marks the variant selected for the final model.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Front and side views of the MHR rest-pose body mesh M0 .
Table 2: Training dataset size comparisons. “N/A” refers to information not publicly available.
Method
Trainable
Backbone
Total
H100 GPU
Training
Params
Params
Params
hours
Hardware
IDOL [ 118 ]
544M
1.2B
1.7B
768
32 × H100
LHM [ 67 ]
1.1B
1.5B
2.6B
3814
64 × A100
LHM++ [ 68 ]
609M
330M
939M
1695
32 × A100
Ours
68M
841M
908M
125
4 × A6000
Appendix
Table 3: Comparison of model size and training compute for models evaluated on NeuMan.
Method
V
Reconstruction
Animation
GPU
(s) ↓
FPS ↑
VRAM ↓
IDOL [ 118 ]
1
2.5
45
17.0
LHM [ 67 ]
1
20.6
85
22.0
LHM++ [ 68 ]
1
2.8
20
7.9
2
4.5
7.9
3
6.6
8.1
Appendix
Table 4: NeuMan runtime profile for IDOL, LHM, LHM++, and ARG-Avatar. Reconstruction time measures feedforward avatar reconstruction latency; animation FPS measures rendering throughput over 104 frames; peak VRAM is the peak GPU memory usage during the profiled run.
Figure 9: Feedforward reconstruction latency on NeuMan for the V=1 IDOL and LHM baselines and the V=6 LHM++ and ARG-Avatar configurations. At six views, ARG-Avatar reconstructs an avatar 6.2× faster than LHM++.
Method
Views
2K2K
THuman2.1
Human4DiT
4D-Dress
DNARendering
HumanVid
Overall
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
SIGMAN
1
15.14
0.905
0.111
15.33
0.910
0.111
15.52
0.912
0.129
14.73
0.897
0.154
16.68
0.947
0.070
16.17
0.862
0.162
15.90
0.897
0.125
IDOL
1
20.88
0.928
0.085
21.05
0.935
0.095
22.15
0.942
0.077
20.63
0.929
0.092
23.44
0.964
0.048
20.53
0.899
0.129
21.25
0.924
0.098
LHM
1
22.55
0.935
0.072
22.81
0.940
0.087
23.17
0.946
0.071
21.65
0.934
0.082
25.75
0.968
0.037
21.23
0.906
0.111
22.38
0.931
0.085
LHM++
1
22.24
0.935
0.070
22.34
0.939
0.086
22.36
0.943
0.074
21.55
0.934
0.079
25.80
0.968
0.036
22.03
0.910
0.104
22.62
0.932
0.082
2
22.23
0.936
0.068
22.50
0.939
0.084
22.52
0.944
0.072
21.53
0.934
0.078
25.78
0.968
0.035
22.35
0.911
0.101
22.82
0.932
0.079
Appendix
Table 5: Per-dataset results on held-out validation splits for novel-pose or novel-view synthesis. Vertical rules separate datasets. “Overall” aggregates results across all evaluated datasets, including NeuMan (reported in the main paper), weighted by the number of cross-samples. First , second , and third indicate the top three results in each metric column.
Figure 10: Ablated ARGBlock variants, drawn with the same visual vocabulary as the full block in Fig. 5 . Naive attention keeps only the embedding and cross-attention path. FACRoPE w/o ITR animates and renders the template rather than a decoded avatar, so FAC guides the cross-attention but no intermediate avatar is decoded.
NeuMan
Multiview Videos
3D Scans
HumanVid
Method
V
Novel Poses
Input Poses
Novel Poses
Novel Views
Novel Poses
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
ARG-Avatar w/ SAM3DBody
1
27.02
0.966
0.024
28.67
0.974
0.016
25.76
0.963
0.034
24.02
0.939
0.059
21.64
0.897
0.100
2
29.17
0.971
0.019
33.88
0.981
0.010
27.35
0.966
0.028
25.34
0.943
0.048
24.57
0.912
0.076
4
30.14
0.973
0.017
33.26
0.981
0.011
28.30
0.969
0.024
26.28
0.946
0.042
25.44
0.916
0.069
6
30.82
0.975
0.016
33.30
0.981
0.011
28.59
0.970
0.023
26.45
0.947
0.041
25.86
0.918
0.067
Appendix
Table 6: Ablation of SAM3DBody and DINOv2 image features.
Figure 11: Additional qualitative comparisons of ARG-Avatar with the baselines on NeuMan. IDOL and LHM use the first view as input; LHM++ and ARG-Avatar use all six views. The first column of each group shows an overlay of the rendered and target images.
Figure 12: Additional qualitative comparisons of ARG-Avatar with the baselines on held-out validation subjects from Human4DiT, THuman2.1, and 2K2K. All samples are unseen during ARG-Avatar training, but 2K2K is used to train LHM and LHM++. Subject names are shown below their respective dataset names.
Figure 13: Additional qualitative comparisons of ARG-Avatar with the baselines on held-out validation subjects from 4D-Dress and DNARendering. Subject names are shown below their respective dataset names.
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.
Thuan Hoang Nguyen, Jiahao Luo, Yinyu Nie +3
Snap Inc. · MBZUAI · University of California, Santa Cruz
We present FFAvatar, a Transformer-based 3D Gaussian framework for fast construction of high-quality and animatable 4D head avatars from one or more reference portrait images. Unlike existing feed-forward approaches that require a fixed number of input views, FFAvatar supports incremental reconstruction, progressively refining the avatar representation as additional reference images become available. At the core of our method is an alternating attention mechanism that disentangles identity appearance from expression and viewpoint variations, enabling the reconstruction of a canonical 3D appearance that remains consistent across poses and facial expressions. To balance visual fidelity and computational efficiency, we introduce a sparse-to-dense learning paradigm. Coarse appearance features are first learned using sparse primitives anchored to the FLAME vertex level and are subsequently densified in the UV domain to capture fine-grained geometric and texture details. We further propose a plug-and-play motion refinement module that enables subject-specific dynamic personalization by modeling residual motion beyond parametric deformation. Extensive experiments demonstrate that FFAvatar efficiently produces high-fidelity and controllable 4D head avatars, achieving superior flexibility, driving efficiency, and identity-consistent rendering across diverse expressions and viewpoints.
Jianjiang Yao, Ke Xian, Renxiang Dai +1
School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China
Recent large human reconstruction models create remarkable animatable 3D avatars from a single image, but their appearance remains entangled with the illumination and cannot be relighted. In contrast, existing 3D relightable-avatar reconstruction methods typically require person-specific optimization. We present ZARA, a zero-shot feed-forward framework that reconstructs an animatable 3D Gaussian avatar together with its intrinsic attributes from a single image. ZARA couples a pretrained large-scale human reconstruction model with a Material Transformer to inherit geometry priors while learning material-aware Gaussian representations. To preserve localized details, we introduce Gaussian-aligned material feature fusion strategy, which injects multi-scale material features onto individual Gaussians and combines them with geometry tokens. As most human-avatar datasets lack ground-truth material maps, we use pseudo-material labels as conditioning inputs and supervision targets during training. We further introduce MatFace regularization to preserve facial structure and color in albedo prediction under unreliable facial pseudo-material supervision. Extensive experiments demonstrate that ZARA reconstructs animatable and relightable humans within tens of seconds without person-specific optimization or post-processing, outperforming existing baselines in both relighting quality and materials generalization ability. As the ability of recovering the physically-based-rendering attributes of the human, ZARA also enables direct material editing of the avatar's appearance.