Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OMG-Avatar: One-shot Multi-LOD Gaussian Head Avatar

Jianqiang Ren, Lin Liu, Steven Hoi

TL;DR

The proposed OMG-Avatar is a novel One-shot method that leverages a Multi-LOD (Level-of-Detail) Gaussian representation for animatable 3D head reconstruction from a single image in 0.2s and introduces a coarse-to-fine learning paradigm to support Level-of-Detail functionality and enhance the perception of hierarchical details.

Abstract

We propose OMG-Avatar, a novel One-shot method that leverages a Multi-LOD (Level-of-Detail) Gaussian representation for animatable 3D head reconstruction from a single image in 0.2s. Our method enables LOD head avatar modeling using a unified model that accommodates diverse hardware capabilities and inference speed requirements. To capture both global and local facial characteristics, we employ a transformer-based architecture for global feature extraction and projection-based sampling for local feature acquisition. These features are effectively fused under the guidance of a depth buffer, ensuring occlusion plausibility. We further introduce a coarse-to-fine learning paradigm to support Level-of-Detail functionality and enhance the perception of hierarchical details. To address the limitations of 3DMMs in modeling non-head regions such as the shoulders, we introduce a multi-region decomposition scheme in which the head and shoulders are predicted separately and then integrated through cross-region combination. Extensive experiments demonstrate that OMG-Avatar outperforms state-of-the-art methods in reconstruction quality, reenactment performance, and computational efficiency.

OMG-Avatar: One-shot Multi-LOD Gaussian Head Avatar

TL;DR

The proposed OMG-Avatar is a novel One-shot method that leverages a Multi-LOD (Level-of-Detail) Gaussian representation for animatable 3D head reconstruction from a single image in 0.2s and introduces a coarse-to-fine learning paradigm to support Level-of-Detail functionality and enhance the perception of hierarchical details.

Abstract

We propose OMG-Avatar, a novel One-shot method that leverages a Multi-LOD (Level-of-Detail) Gaussian representation for animatable 3D head reconstruction from a single image in 0.2s. Our method enables LOD head avatar modeling using a unified model that accommodates diverse hardware capabilities and inference speed requirements. To capture both global and local facial characteristics, we employ a transformer-based architecture for global feature extraction and projection-based sampling for local feature acquisition. These features are effectively fused under the guidance of a depth buffer, ensuring occlusion plausibility. We further introduce a coarse-to-fine learning paradigm to support Level-of-Detail functionality and enhance the perception of hierarchical details. To address the limitations of 3DMMs in modeling non-head regions such as the shoulders, we introduce a multi-region decomposition scheme in which the head and shoulders are predicted separately and then integrated through cross-region combination. Extensive experiments demonstrate that OMG-Avatar outperforms state-of-the-art methods in reconstruction quality, reenactment performance, and computational efficiency.
Paper Structure (22 sections, 13 equations, 12 figures, 6 tables)

This paper contains 22 sections, 13 equations, 12 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Hierarchical Global-Local Feature Extraction
  5. Occlusion-Aware Feature Fusion
  6. Multi-region Modeling and Integration
  7. Reenactment and Rendering
  8. Learning Strategy
  9. Experiment
  10. Datasets and Settings
  11. Baseline methods
  12. Qualitative Evaluation
  13. Quantitative Evaluation
  14. Ablation Studies
  15. Conclusion
  16. ...and 7 more sections

Figures (12)

  • Figure 1: The overall pipeline of OMG-Avatar framework. Our method extracts global features via cross-attention and local details via projection-based sampling, which are fused under the guidance of depth buffers. A coarse-to-fine strategy is proposed to facilitate hierarchical detail perception. The head and shoulder are predicted separately using shared features and then combined for rendering.
  • Figure 2: Since the original FLAME model lacks shoulder regions and sufficient geometric representation for high-resolution LOD requirements, we progressively subdivide the head mesh during training and enhance it with predicted shoulder geometry through cross-region combination. Sub #$i$ indicates that the head mesh has been subdivided $i$ times.
  • Figure 3: Cross-identity reenactment results on VFHQ and HDTF datasets.
  • Figure 4: Our local-global feature fusion (OAFF) and multi-region fusion strategy significantly improve identity consistency and completeness in non-head regions. The neural refiner further boosts visual fidelity, especially for dynamic facial expressions.
  • Figure 5: The correlation between Gaussian count and reconstruction performance on the VFHQ dataset.
  • ...and 7 more figures