Given a 3D static avatar from any multimodal input, e.g. monocular image with LHM, we introduce a novel method, AvatarDynamizer, which converts static avatars into controllable 4D human avatars with pose-dependent surface dynamics (see the shirt).
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.
Given a static human avatar from multimodal inputs, e.g., multi-view images, our method first perform identity fitting to obtain coarse shape and identity textures.Our Dynamic Texture Generator takes identity and pose information to predict dynamic texture maps. Then, our Generalized Gaussian Decoder recovers Gaussian splats and outputs faithful and view-consistent 4D avatar renderings.
a)Regression-based approaches are effective for static 3D reconstruction, but their extension to dynamic humans remains difficult. Deterministic regression is not well suited for modeling fine-grained surface dynamics under limited data, as motion-dependent details such as wrinkles are high-frequency and inherently ambiguous: the same identity and pose can correspond to multiple plausible surface states depending on unobserved physical factors. As a result, these methods often produce over-smoothed or nearly static dynamics, even when motion conditions are provided. To alleviate this limitation, they typically require large-scale multi-view motion data, which is expensive to collect and often not publicly. b) Another option is to condition a video model on a static human rendering, allowing the model to generate plausible dynamics from learned 2D video priors. However, since the generation is performed in image space, the results do not explicitly guarantee 3D consistency across time and viewpoints. c) Our representation is designed to address this gap. It leverages video priors for dynamic generation while using the Generalized Gaussian Decoder to produce a dynamic 3D human representation that remains consistent in 3D.
Existing datasets fall into two extremes: DDC and DUT provide rich clothing dynamics but cover very few identities. In contrast, MVHumanNet++ and DNA-Rendering include many subjects, but these subjects mostly perform slow or static motions. Our DynaHuman bridges this gap, featuring both scale (58 subjects) and dynamic diversity, enabling generalizable learning of realistic cloth dynamics.
[1] Real-time Deep Dynamic Characters (Habermann et al. 2022)
[2] Real-time Free-view Human Rendering from Sparse-view RGB Videos using Double Unprojected Textures (Sun et al. 2025)
[3] Implicit Neural Representations with Structured Latent Codes for Novel View Synthesis of Dynamic Humans (Peng et al. 2021)
[4] A Diverse Neural Actor Repository for High-Fidelity Human-centric Rendering (Cheng et al. 2023)
[5] A Large-scale Dataset of Multi-view Daily Dressing Human Captures with Richer Annotations for 3D Human Digitization (Li et al. 2025)
Although trained with a larger model and more data, 2D-only method Wan-Animate-2 shares similar consistency issues with GAS. Besides, it could fail for some pose control. Our method produces temporally consistent and faithful wrinkles and maintains consistency under novel-view rendering.
[6] Wan-Animate-2: Pushing the Application Boundaries of Character Animation (Tongyi Lab 2026)
Our method transforms static avatars that are acquired from multi-modal inputs, i.e texts and 3D mesh scans, into dynamic ones.
We use LHM to create static avatars for in-the-wild monocular images and drive them with the same body motion under dynamic camera views. Notice that how the wrinkles evolve with the motion and vary across different clothing types.
[7] LHM: Large Animatable Human Reconstruction Model for Single Image to 3D in Seconds (Qiu et al. 2025)
@article{sun2026avatardynamizer,
title = {AvatarDynamizer: From Static to Dynamic Human Avatars via Generative Dynamic Textures},
author = {Sun, Guoxing and Zhu, Heming and Lyu, Linjie and Fua, Pascal and Theobalt, Christian and Habermann, Marc},
year = {2026},
eprint={2608.19900},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2608.19900},
}