CloseUpAvatar: High-Fidelity Animatable Full-Body Avatars with Mixture of Multi-Scale Textures

David Svitov, Pietro Morerio, Lourdes Agapito, Alessio Del BuearXiv 2025 · DOI · arXiv

Abstract

CloseUpAvatar is a novel representation for animatable full-body avatars designed to maintain high-fidelity details during extreme close-ups. By utilizing "textured surfels" combined with a MipMapping-inspired Mixture of Multi-Scale Textures (MST) blending approach, the method achieves state-of-the-art perceptual quality and real-time rendering speeds (up to 350 FPS). This approach effectively addresses the "scale dilemma" inherent in current full-body avatar models, where zooming in often reveals blurry artifacts or primitive shapes. CloseUpAvatar requires significantly fewer primitives than previous methods while delivering superior visual sharpness for immersive environments.

Introduction: The Scale Dilemma in Digital Humans

Creating realistic digital humans is a cornerstone of VR, telepresence, and gaming. However, current full-body avatar methods face a significant "scale dilemma." Representations like 3D Gaussian Splatting (3DGS) often look excellent from a distance but reveal blurry artifacts or visible "splat" shapes when a user zooms in—a common occurrence in immersive VR environments.

While specialized head-only avatars can handle close-ups, full-body models struggle to balance the high-frequency details needed for facial features or fabric textures with the efficiency required to render an entire moving body in real-time. Existing methods are often not explicitly designed to maintain consistency across varying camera distances, leading to a loss of realism during transitions from far-away views to close-up interactions.

Methodology: Mixture of Multi-Scale Textures (MST)

The core innovation of CloseUpAvatar is the Mixture of Multi-Scale Textures (MST) applied to textured surfels (also known as billboards).

Overview of the CloseUpAvatar pipeline, showing the use of textured surfels and the blending of low and high-frequency textures (MST).
  1. Textured Surfels: Instead of using hundreds of thousands of simple point-like Gaussians, the system uses a smaller set (approx. 20k) of 2D planes (surfels). Each surfel covers a continuous area of the body and carries learnable textures for color and opacity.
  2. Multi-Scale Textures: Each surfel stores two sets of textures: a low-frequency texture ($MST^L$, $16 \times 16$) for general appearance and a high-frequency texture ($MST^H$, $32 \times 32$) that acts as an offset to add sharp details.
  3. Dynamic Blending: Inspired by classic computer graphics MipMapping, the system calculates a view-dependent coefficient based on the surfel's size on the screen. As the camera moves closer, the high-frequency texture is gradually blended in to maintain sharpness.
  4. Camera Augmentation: To train the model for these scales, the authors use a data augmentation strategy, cropping and padding high-resolution (4K) training images to simulate extreme close-ups and far-away views.

Experimental Results

CloseUpAvatar was evaluated using the ActorsHQ dataset (12MP video) and motion sequences from AMASS. It outperforms state-of-the-art baselines like Mmlphuman and AnimatableGaussians in both visual quality and efficiency.

Comparison of CloseUpAvatar against a baseline during zoom-in. Our method maintains higher PSNR and SSIM while achieving superior real-time frame rates.
  • Perceptual Quality: In "zoom-in" scenarios, the model achieved a significantly better LPIPS score (0.2231) compared to Mmlphuman (0.3004), indicating much higher perceived sharpness. Visual results show much clearer eyes, skin pores, and clothing textures.
  • Speed and Efficiency: The method maintains real-time performance, reaching 244 FPS for close-ups and up to 350 FPS for full-body views on an NVIDIA RTX 4090. Furthermore, it is faster to train, requiring only 9 hours compared to the 11–45 hours required by competing methods.
  • Quality over Quantity: By using richer primitives (textured billboards) instead of more numerous but simpler ones, the model achieves better detail with 10x fewer primitives (20k vs. 200k).

Conclusion and Future Work

CloseUpAvatar demonstrates that adapting classical computer graphics techniques like MipMapping to modern neural rendering can robustly handle level-of-detail (LOD) challenges. While highly effective, the method currently faces limitations with highly complex non-rigid deformations (like loose clothing) and does not yet implement specialized models for fine finger movements or subtle facial expressions like blinking. Future iterations aim to explore more sophisticated animation for these high-detail areas and test performance on lower-quality monocular video inputs.

Summary updated

BibTeX
@misc{svitov2025closeupavatar,
  author = {David Svitov and Pietro Morerio and Lourdes Agapito and Alessio Del Bue},
  title = {{CloseUpAvatar: High-Fidelity Animatable Full-Body Avatars with Mixture of Multi-Scale Textures}},
  year = {2025},
  eprint = {2512.03593},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.03593},
}
Figure