cs.CVAug 20, 2026
SaveAvatarDynamizer: From Static to Dynamic Human Avatars via Generative Dynamic Textures
Organizations: Max Planck Institute for Informatics, Saarland Informatics Campus · EPFL · VIA Research Center
Abstract
For full-body avatars, modeling surface dynamics is crucial for overcoming the uncanny valley and achieving perceptual realism. Person-agnostic methods recover static 3D avatars from monocular images, videos, or text prompts, but their skeleton-driven animations lack realistic surface dynamics such as clothing wrinkles. In contrast, person-specific methods achieve high-quality rendering and realistic dynamics, but require expensive multi-view captures for each individual. Recent generalizable dynamic avatar methods struggle to embed surface dynamics, leading to either limited multi-view consistency or dynamic expressiveness. To this end, we propose AvatarDynamizer, a generative method that transforms an off-the-shelf static 3D avatar into a controllable, realistic, and multi-view-consistent 4D avatar. We introduce a novel texture-space surface-dynamics embedding and formulate avatar dynamics modeling as conditional texture generation. Our encoder--decoder representation embeds pose-dependent dynamics into dynamic texture maps, enabling compatibility with pre-trained video diffusion models while decoding them into 3D Gaussians for multi-view consistent rendering. Since existing datasets are limited in scale, sequence length, or motion diversity, we collect a large-scale multi-view dataset with long sequences covering diverse skeletal motions and surface dynamics. Experiments show that our method effectively animates static avatars with faithful surface dynamics and outperforms competing generalizable methods in visual fidelity, especially under limited dynamic training data.
Figures & tables
Figure 2: Overview . Given a static human avatar reconstructed from multimodal inputs, e.g., multi-view images, we first perform identity fitting to obtain its coarse shape and identity texture. Our Dynamic Texture Generator predicts dynamic texture maps conditioned on identity and novel pose. Then, our Generalized Gaussian Decoder recovers Gaussian splats for faithful and view-consistent renderings.
Figure 3: Qualitative Comparison on DynaHuman . Given a static avatar reconstructed from a monocular image using LHM [ 59 ] or from multi-view images (MV), we compare our method with state-of-the-art generalizable surface dynamics methods. Our AvatarDynamizer produces faithful wrinkles while preserving both 3D geometry and identity consistency.
| Dataset | Method | PSNR | SSIM | LPIPS | FID | FVD | Method | PSNR | SSIM | LPIPS | FID | FVD |
| DeepCap | LHM | 30.61 | 0.715 | 0.201 | 50.10 | 150.12 | MV | 30.86 | 0.721 | 0.191 | 39.27 | 142.18 |
| +GAS | 30.14 | 0.700 | 0.248 | 57.14 | 223.82 | +GAS | 30.48 | 0.705 | 0.242 | 54.28 | 222.78 | |
| +VAP | 31.31 | 0.730 | 0.221 | 101.73 | 206.88 | +VAP | 32.64 | 0.752 | 0.197 | 73.47 | 178.08 | |
| +Ours | 31.19 | 0.723 | 0.200 | 48.69 | 160.75 | +Ours | 31.95 | 0.735 | 0.185 | 37.05 | 134.26 | |
| DDC | LHM | 29.65 | 0.696 | 0.206 | 42.64 | 108.00 | MV | 29.52 | 0.698 | 0.189 | 36.14 | 103.10 |
| +GAS | 29.40 | 0.676 | 0.281 | 44.32 | 167.20 | +GAS | 29.64 | 0.680 | 0.279 | 44.35 | 164.61 |
Table 1: Quantitative comparison across datasets. Each result is averaged over two subjects. Higher PSNR and SSIM are better, while lower LPIPS, FID, and FVD are better. The best and second-best results are highlighted within each dataset and each method family.
| Representation | PSNR | SSIM | LPIPS | Tex/Full Storage |
| ST + Decoder (VAP) | 30.36 | 0.808 | 0.160 | 89KB / 67.6GB |
| RDT + Decoder (Ours) | 32.02 | 0.874 | 0.103 | 110MB / 67.6GB |
Table 2: Quantitative Comparison on Human Representation. We compare our human representation with VAP [ 21 ] on Subject59 from the training set. Our Generalized Gaussian Decoder (GGD) with optimized reference dynamic textures (RT) achieves better image reconstruction with only a moderate storage increase.
| Train Set | Test Set | PSNR | LPIPS | FVD |
| w/o DynaHuman | DynaHuman | 28.77 | 0.142 | 33.60 |
| w/o MVHPP | DynaHuman | 29.70 | 0.124 | 25.77 |
| Ours | DynaHuman | 29.71 | 0.122 | 22.97 |
| w/o DynaHuman | MVHPP | 29.45 | 0.190 | - |
| w/o MVHPP | MVHPP | 27.22 | 0.213 | - |
| Ours | MVHPP | 29.51 | 0.190 | - |
Table 3: Quantitative Ablation on Training Datasets. Training with both datasets results in improved generalization and higher reconstruction accuracy.
| Baselines | PSNR | SSIM | LPIPS | FID | FVD |
| w/o PP | 24.70 | 0.700 | 0.240 | 60.42 | 228.21 |
| w/o TI | 26.61 | 0.748 | 0.187 | 16.91 | 55.20 |
| Ours | 26.61 | 0.745 | 0.182 | 13.84 | 37.74 |
| w/ RDT | 29.89 | 0.847 | 0.116 | 8.47 | 20.73 |
Table 4: Quantitative Ablation on Dynamics Prior with Subject71. Without pretrained diffusion weights, our Dynamic Texture Generator has limited performance. Removing the temporal block results in degraded temporal performance.
Figure 4: Qualitative Ablation on Dynamics Prior. Training the Dynamic Texture Generator from scratch without pretrained diffusion weights yields limited quality. Removing the temporal block reduces temporal consistency and causes frame-to-frame variations (red arrows). The full model achieves the best generation quality and temporal consistency.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Multi-modal Avatar Generation . Our method transforms static avatars that are acquired from multi-modal inputs into dynamic ones. Here, we show static avatars that are created from text and 3D scans. Next, the static avatar is converted into identity texture and coarse shape. Finally, we can drive the avatar using novel skeletal motions. We highlight the different wrinkle patters and dynamics that are induced by the motion and that our method can faithfully generate.
Figure 6: Conceptual Comparison of Generalized and Dynamic Human Representations . a) Regressing 3D human dynamics directly cannot leverage foundational video priors. Thus, such methods rely on large-scale (closed-source) multi-view data. While regression-based models perform well for static reconstruction, they struggle with dynamic reconstruction. b) Leveraging a static rendering as conditioning for a video model allows generative modeling of dynamics. However, since the model operates in 2D image space, 3D consistency is not ensured. c) Our human representation unites the best of both worlds and enables generative modeling and video prior usage for dynamics while our Generalized Gaussian Decoder predicts a dynamic 3D human that is 3D consistent.
| Dataset | Frames | Views | Identities | Motion | Geometry | Texture | 4K | Frames |
| Per Subject | Diversity | Resolution | (In total) | |||||
| DDC [ 23 ] | 90 | 5 | ✓ | ✗ | ✗ | ✓ | 11.8M | |
| DUT [ 70 ] | 116 | 2 | ✓ | ✓ | ✗ | ✓ | 8.1M | |
| ZJU [ 57 ] | 500- | 24 | 9 | ✗ | ✗ | ✗ | ✗ | 180K |
| DNA [ 12 ] | 200 | 60 | 500 | ✗ | ✓ | ✗ | ( ✓ ) | 67.5M |
| MVHPP [ 44 ] | 60 | 16 | ✗ | ✓ | ✗ | ✗ | 645.1M |
Table 5: Dataset Comparison. Statistics are based on the publicly available versions of the datasets (e.g., MVHPP [ 44 ] only releases part of the data). Our dataset matches or surpasses person-specific approaches [ 56 , 23 , 70 ] in motion diversity and number of views while providing significantly more identities. Compared to large-scale identity datasets [ 12 , 44 ] , our dataset has significantly higher motion diversity, views, and frames per subject. Besides, we provide high-precision geometry and registered textures. Thus, our dataset bridges an important gap in literature to cover surface dynamics.
Figure 7: Representation Ablation. The qualitative ablation over the avatar representation on the training split. Given the same multi-view videos, compared with the baseline approach, i.e., VAP [ 21 ] , our proposed Generalized Gaussian Decoder (GGD), which takes the reference dynamic texture (RDT) as input, reconstructs the pose-dependent wrinkles more accurately.
Figure 8: Dataset Ablation. The qualitative ablation over the training datasets. Removing the DynaHuman dataset from training leads to uneven lighting, while excluding the MVHPP dataset [ 44 ] from training could cause severe artifacts. By combining our DynaHuman dataset and the MVHPP dataset, our Generalized Gaussian Decoder achieves better reconstruction.
Figure 9: Dataset Visualization . For each subject, we provide an A-pose 3D scan, multi-view videos, and registered and deformed SMPL-X as well as dynamic texture annotations.
| Identity Consistency (Q1) | Physical Plausibility (Q2) | View Consistency (Q3) | Overall Score (Q4) | |
| Static Avatar | 3.944/5.0 | 3.289/5.0 | 4.089/5.0 | 3.644/5.0 |
| +GAS | 1.233/5.0 | 1.289/5.0 | 1.356/5.0 | 1.256/5.0 |
| +VAP | 3.511/5.0 | 2.622/5.0 | 3.689/5.0 | 2.878/5.0 |
| +Ours | 4.000/5.0 | 4.003/5.0 | 3.789/5.0 | 3.947/5.0 |
Table 6: User study questionnaire results. 1 = Poor, 5 = Excellent.
Figure 10: Failure of Wan-Animate 2 . State-of-the-art 2D-only animation models fail to preserve spatial and temporal consistency. Note how the back texture and nail colors change over time (temporal inconsistency), and how facial wrinkles, clothing wrinkles, and nail shapes change across different camera views (spatial inconsistency). All results shown above are from the official Wan-Animate 2 results.
Figure 11: Additional Failure of Wan-Animate 2 . Given a DynaHuman subject, Wan-Animate 2 exhibits several limitations, including incorrect motion embedding, pose-to-appearance ambiguity, inconsistent lighting, distorted body shape, and background artifacts.
| Subject71 | Subject76 | |||
| FID | FVD | FID | FVD | |
| Wan-Animate 2 | 31.83 | 185.09 | 28.20 | 179.98 |
| LHM+Ours | 28.20 | 72.66 | 19.65 | 40.78 |
Table 7: Quantitative comparison on DynaHuman. Here, we compare Wan-Animate 2 [ 77 ] with our method. Our method achieves consistent better results.
Figure 12: Qualitative Comparisons on DeepCap, DDC and MetaCap . We show additional qualitative comparisons on public datasets, i.e. DeepCap [ 22 ] , DDC [ 23 ] and MetaCap [ 69 ] . The results are shown in the order of ground truth , MV (static avatar) , and our method . Our method successfully generates pose-dependent wrinkles across different motions and clothing types.
Figure 13: Qualitative Comparisons on In-the-wild Images . The leftmost image is the input to LHM [ 59 ] . The results are shown in the order of LHM (static avatar) , and our method . We drive them with the same body motion. Notice that how the wrinkles evolve with the motion and vary across different clothing types.’