OMG-Avatar: One-shot Multi-LOD Gaussian Head Avatar
About
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. The project homepage is https://human3daigc.github.io/OMGAvatar_project_page/ .
Key insights
What problem does OMG-Avatar solve?
It reconstructs animatable 3D head avatars from a single image in 0.2s using multi-LOD Gaussians, enabling dynamic detail levels for diverse hardware. It fuses global and local features with depth-guided occlusion handling and separately models shoulders.
How does it differ from prior work?
Unlike LAM's attention over roughly 80,000 vertices, OMG-Avatar applies attention at a coarse 5,000-vertex level before refinement. Its 29,000-Gaussian variant surpasses LAM and GAGAvatar on the reported reconstruction benchmarks with fewer Gaussian primitives.
What are the practical uses?
It targets video production, virtual meetings, and Metaverse applications, achieving 85 FPS on A100 and 126 FPS on RTX 4090. These are benchmark-prototype results; the paper documents no production deployment.
What limitations do the authors note?
FLAME cannot capture fine dynamics like tongue motion or hair deformation, limiting expressiveness. Training on monocular videos reduces robustness beyond ±60° head rotation; multi-view data is planned to address this.