One Avatar, Any Budget:
Robust LoD for Dynamic Gaussian Avatars

NeurIPS 2026 (Poster)

Haoyu Zhao, Wenjuan Gong, Junang Qi, Haiyu Zhang, Wenjian Zhao, Zimu YOU

RobustAvatar adapts a single dynamic Gaussian avatar to different rendering budgets through semantic-aware selection and canonical compensation, without ratio-specific fine-tuning.

Abstract

Cross-device deployment of a single reconstructed avatar often requires adaptation to varying Gaussian budgets with minimal degradation in visual quality. Motivated by the need for a robust budget-adaptive 3D representation, we propose RobustAvatar, a train-once Level-of-Detail (LoD) adaptation framework for dynamic Gaussian avatars based on sequential selection and compensation. RobustAvatar supports flexible compression-ratio control from a single trained avatar, without requiring ratio-specific fine-tuning or the maintenance of separate LoD models. The framework consists of two coordinated modules: Semantic Distribution-Regularized Selection (SDS) and Canonical Budget-Conditioned Compensation (CBC). SDS performs budget-aware Gaussian selection guided by semantic distribution regularization, preserving the coverage of perceptually salient facial regions in the full avatar under constrained budgets. CBC refines the selected primitives in a canonical, expression- and pose-neutral space by predicting ratio-conditioned attribute residuals, reducing detail loss from Gaussian removal and improving robustness to decreasing Gaussian budgets. Experiments on the NeRSemble dataset show that RobustAvatar achieves more robust budget adaptation than state-of-the-art methods, with slower quality degradation as Gaussian budgets decrease. The code will be made publicly available.

Read the complete paper

Method Overview

Overview of RobustAvatar. Given a full Gaussian avatar and a target budget ratio r, Semantic Distribution-Regularized Selection (SDS) selects the top-K primitives while preserving the distribution of important facial regions. Canonical Budget-Conditioned Compensation (CBC) then predicts ratio-conditioned attribute residuals in an expression- and pose-neutral space. The compensated subset is transformed using FLAME and rasterized, enabling flexible budget control with a single trained model.

Results at Different Gaussian Budgets

RobustAvatar preserves identity and expression across different retained-Gaussian ratios. Select a budget and subject to compare the rendered avatar with the full-budget result and ground truth.

Retained Gaussian budget
RobustAvatar10% budget
Full budget100%
Ground truthReference

Qualitative results from Figure 4. Colored boxes show local facial details. Click any image to enlarge.

Comparisons against ArchitectHead

Under comparable Gaussian budgets, RobustAvatar preserves sharper expression-sensitive details around the eyes, lips, teeth, and ears.

ArchitectHead8,788 Gaussians
RobustAvatar8,000 Gaussians
Ground truthReference

Figure 3. Gaussian counts follow the figure labels; Table 1 reports aggregate counts for the quantitative evaluation.

Quantitative Evaluation

We evaluate novel view synthesis (NVS) and novel expression synthesis (NES) on NeRSemble. Gaussian counts are shown directly because the methods use different ratio definitions.

PSNR vs. Gaussian count

Novel view synthesis · Higher PSNR is better

RobustAvatarArchitectHeadProgressiveAvatars*Full GaussianAvatars

* ProgressiveAvatars values are cited results, not a reproduction. Lines connect the reported settings; no intermediate measurements are implied.

View the reported measurements
Selected PSNR measurements from Table 1
Method / settingGaussiansNVS PSNR ↑NES PSNR ↑
RobustAvatar (5%)2,30726.3624.17
RobustAvatar (10%)4,61528.4225.05
RobustAvatar (20%)9,23029.4725.54
RobustAvatar (50%)23,07629.7225.67
ArchitectHead (0.90)4,87126.1817.83
ArchitectHead (0.75)8,78826.6218.75
ArchitectHead (0.45)23,13327.7218.72
ProgressiveAvatars (5%)*10,14427.8925.13
ProgressiveAvatars (25%)*37,30229.1425.58
GaussianAvatars (100%)46,15229.1925.55

BibTeX

Provisional citation for the anonymous manuscript. Author and publication metadata will be updated with the final release.

@misc{robustavatar2026,
  title = {One Avatar, Any Budget: Robust LoD for
           Dynamic Gaussian Avatars},
  author = {Zhao, Haoyu and Gong, Wenjuan and Qi, Junang and Zhang, Haiyu and Zhao, Wenjian and YOU, Zimu},
  year = {2026},
  note = {Accepted at NeurIPS 2026}
}