Research map / 3D vision / Face and shape
Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Gaussian Head Avatar, an animatable head avatar built from expression-conditioned dynamic 3D Gaussians with a fully learned MLP deformation field, initialized via an SDF/DMTet geometry guidance model, and rendered at 2K with a super-resolution network.
Problem
High-fidelity, animatable 3D head avatar reconstruction at 2K resolution from lightweight sparse multi-view video, where NeRF-based methods lack pixel-level detail and LBS/FLAME-based deformation cannot capture exaggerated expressions.
Approach
Replace LBS/mesh-based deformation with a fully learned, expression- and pose-conditioned MLP that displaces 3D Gaussians (decoupled by distance to 3D landmarks), and initialize the Gaussians from a geometry-guided SDF+DMTet mesh model whose deformation and color MLPs are carried over to training.
Key results
- Self-reenactment at 2K: FID 18.50 and LPIPS 0.098 versus 31.06 and 0.209 for HAvatar, with PSNR 27.70 vs 27.19.
- Novel view synthesis (8 training views) reports PSNR 27.58 and LPIPS 0.059, outperforming HAvatar (27.13 PSNR, 0.65 LPIPS as printed).
- Ablation: geometry-guided init improves PSNR to 28.94 over FLAME-Init at 28.73 and Mesh-Deform at 28.83.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Gaussian Head Avatar instead of HAvatar | head avatar reconstruction · Self reenactment | LPIPS (2K): 0.209 -> 0.098 | Helped |
| Use Gaussian Head Avatar instead of HAvatar | head avatar reconstruction · Self reenactment | FID (2K): 31.06 -> 18.50 | Helped |
| Use Gaussian Head Avatar instead of HAvatar | head avatar reconstruction · Self reenactment | PSNR: 27.19 -> 27.70 | Helped |
| Use Gaussian Head Avatar instead of HAvatar | head avatar reconstruction · Novel view synthesis (3D consistency) | PSNR: 27.13 -> 27.58 | Helped |
| Use Gaussian Head Avatar instead of HAvatar | head avatar reconstruction · Novel view synthesis (3D consistency) | LPIPS: 0.65 -> 0.059 | Helped |
| Add the super-resolution network (ablation) | Gaussian Head Avatar · Self reenactment | LPIPS (2K): 0.202 -> 0.098 | Helped |
| Add the super-resolution network (ablation) | Gaussian Head Avatar · Self reenactment | PSNR: 27.82 -> 27.70 | Hurt |
| Add geometry-guided initialization instead of FLAME initialization (ablation) | Gaussian Head Avatar · Self reenactment | LPIPS: 0.123 -> 0.108 | Helped |
| Add the fully learned deformation field instead of mesh-based deformation (ablation) | Gaussian Head Avatar · Self reenactment | LPIPS: 0.116 -> 0.108 | Helped |
Surprising findings
- Without super-resolution the method's PSNR (27.82) is slightly higher than the full model (27.70), while SR substantially improves LPIPS and FID at 2K.
Limitations
- Blurring for tongue, teeth inside the mouth, and long hair due to lack of tracking.
- Requires multi-view video with 3DMM/landmark preprocessing and background removal.
- Training uses lightweight sparse views (8-16 cameras) but evaluation is limited to a small number of identities.
Experimental setup
Datasets: NeRSemble HAvatar multi-view videos
Benchmarks: NeRSemble HAvatar multi-view videos
Metrics: PSNR SSIM LPIPS FID
Baselines: NeRFBlendShape NeRFace HAvatar FLAME-Init Mesh-Deform
Builds on: 3D Gaussian splatting DMTet NeRSemble HAvatar VGG perceptual loss
Closest papers by meaning
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.94)
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · similarity 0.92)
- 3D Gaussian Head Avatars with Expressive Dynamic Appearances by Compact Tensorial Representations (CVPR 2025 · similarity 0.92)
- Generalizable and Animatable Gaussian Head Avatar (NeurIPS 2024 · similarity 0.91)
- MeGA: Hybrid Mesh-Gaussian Head Avatar for High-Fidelity Rendering and Head Editing (CVPR 2025 · similarity 0.91)
- Gaussian Eigen Models for Human Heads (CVPR 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.