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.
Figures & tables
Method
Zero- Shot
Relightable
PBR
Single-image Relighting
DNF [ 20 ]
×
✓
✓
×
SG-IA [ 21 ]
×
✓
✓
×
IDOL [ 12 ]
✓
×
×
–
LHM [ 11 ]
✓
×
×
–
S3F [ 10 ]
✓
✓
×
✓
LCA [ 14 ]
✓
✓
✓
N/R
TABLE I: Comparison of existing avatar reconstruction methods and the proposed ZARA. Each column indicates whether the method reconstructs avatars in a zero-shot manner, whether it supports relighting, whether it generates PBR materials, and whether it can be relighted form single image input.
Fig. 1: Overview of ZARA and its core network components. Given a single image, the pretrained geometry branch reconstructs canonical Gaussian geometry, while the material branch encodes material maps and predicts per-Gaussian intrinsic attributes. The Gaussian decoding branches predict geometry, pose-dependent geometry offsets, and materials. The material encoder receives the prepared material priors and extracts features that are fused into global material tokens and per-Gaussian material features. The resulting canonical avatar is animated by the target motion and rendered under novel illumination.
Fig. 2: Details of the material encoder. A seven-channel material prior is processed by four ConvNext-style multi-scale encoding blocks, whose features are fused into global material tokens and per-Gaussian material tokens.
Fig. 3: Schematic of MatFace Regularization. The captured RGB face supplies facial structure, while pseudo-label albedo supplies the intrinsic color palette. Their complementary cues supervise the predicted albedo face region during training.
Human4DiT-3D [ 60 ]
RANA [ 38 ]
Method
Albedo
Normal
Albedo
Normal
PSNR ↑
SSIM ↑
LPIPS ↓
MAE( ∘ ) ↓
PSNR ↑
SSIM ↑
LPIPS ↓
MAE( ∘ ) ↓
Available data: full subject-specific video
IntrinsicAvatar † [ 61 ]
21.853
0.8362
0.0845
9.521
22.696
0.8730
0.0541
10.781
DNF-Avatar † [ 20 ]
15.439
0.7003
0.0738
17.785
22.243
0.8647
0.0567
10.508
Available data: a single image
TABLE II: Quantitative comparison of material prediction on Human4DiT-3D and RANA. Person-specific methods ( † ) use full subject-specific videos and iterative optimization while zero-shot methods uses a single image without test-time optimization. For the zero-shot baselines, we report albedo metrics and normal predictions are unavailable. The best result for each metric is bolded separately within each group.
Human4DiT-3D [ 60 ]
RANA [ 38 ]
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FC ↓
PSNR ↑
SSIM ↑
LPIPS ↓
FC ↓
Runtime ↓
Available data: full subject-specific video
IntrinsicAvatar † [ 61 ]
20.373
0.8276
0.0546
0.0461
18.332
0.8693
0.0367
0.0368
2.3 h
DNF-Avatar † [ 20 ]
17.372
0.7441
0.0534
0.0474
18.935
0.8715
0.0510
0.0363
6.6 h
Available data: a single image
LHM [ 11 ] + UniRelight [ 18 ]
14.949
0.6330
0.1353
0.0892
15.413
0.6636
0.2040
0.0974
101.70 s
TABLE III: Quantitative comparison of relighting RGB on Human4DiT-3D and RANA. Person-specific methods ( † ) use full subject-specific videos and iterative optimization, whereas zero-shot methods use a single image. The best result for each metric is bolded separately within each group.
Fig. 4: Qualitative comparison between ZARA, zero-shot baseline methods and person-specific methods with in-the-wild video (the first two samples), studio-captured video (third sample), and synthesis data (the last two samples). For each sample, the first row shows the input view and the second row shows the unseen view. Ref, Pose, Env, A, R, N, Re, and Rec are the reference image, target pose, environment map for relighting, albedo, roughness, normal, relighting rgb, and reconstructed rgb, respectively. For the two-stage zero-shot baselines, we show reconstructed RGB, estimated albedo, and relighting results. For ZARA and person-specific approaches, we visualize the material maps and relighting outputs. Below the dashed line at the bottom are references ground truth from the synthetic dataset, including material maps and RGB images.
Fig. 5: Qualitative ablation of Gaussian-aligned material fusion on two real-world samples. For each sample, the reference image and target pose are followed by RGB/Relit, albedo, and normal results from the full model and the ablated model. The orange boxes indicate localized garment regions used for comparison.
Fig. 6: Qualitative ablation of MatFace Regularization. From left to right, we show the reference image, albedo and relighting results without MatFace, and the corresponding full-model results. Red boxes indicate the facial regions and their enlarged views.
Fig. 7: Qualitative ablation of explicit material conditioning. For each sample, the reference image and target pose are followed by intrinsic-material and relighting results from the full model and the RGB-only variant without the explicit material-conditioning pathway.
Albedo
Normal
Relighting RGB
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
MAE( ∘ ) ↓
PSNR ↑
SSIM ↑
LPIPS ↓
FC ↓
w/o Gaussian-aligned Proj.
20.766
0.8613
0.0665
18.681
18.232
0.8236
0.0781
0.0539
w/o MatFace Regularization
21.430
0.8640
0.0530
14.926
19.506
0.8467
0.0578
0.0826
w/o Explicit Material Cond.
18.316
0.8256
0.0948
21.508
16.475
0.7929
0.1010
0.0572
Full model
21.753
0.8734
0.0509
14.334
19.609
0.8574
0.0539
0.0464
TABLE IV: Quantitative ablation on Human4DiT-3D and RANA. Each metric is averaged over the two datasets and the best metrics are bolded.
Fig. 8: Downstream material editing applications. Top: Albedo editing adds a logo and changes garment colors, with the corresponding relighting results shown below. Bottom: Normal editing changes the distribution of garment folds and the resulting relighting appearance.
Fig. 9: Failure-case under corrupted pseudo-material input. The upper block shows the reference RGB image and erroneous pseudo-albedo, pseudo-roughness, and pseudo-normal maps, while the lower block presents ZARA’s relighting appearance and predicted intrinsic materials.
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
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.
Reconstructing animatable 3D human avatars from a single image remains particularly challenging for loose garments, whose geometry and motion cannot be adequately represented by body-aligned topology and skinning. We present Forwardrobe, a feed-forward framework for reconstructing garment-aware Gaussian avatars from a single image. Forwardrobe explicitly separates clothing from the body in canonical Gaussian space and equips the garment layer with continuity-aware geometry and skinning initialization, pose-conditioned non-rigid deformation, and appearance adaptation. These designs improve garment reconstruction and visual quality during animation, particularly for skirts and dresses. The separated garment layer additionally forms an independently controllable 3D asset, enabling garment editing, transfer, and 3D virtual try-on. Experiments demonstrate improved garment reconstruction quality and greater flexibility in garment manipulation compared with existing single-image avatar reconstruction methods.
Daisheng Jin, Shuyun Wang, Ying He
Nanyang Technological University · Singapore · The University of Queensland +1