Research map / 3D vision / Face and shape
GaussianAvatars: Photorealistic Head Avatars with Rigged 3D 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.
Rigs 3D Gaussian splats to FLAME mesh triangles, with binding inheritance for densification and position/scaling regularization, to build photorealistic animatable head avatars from multi-view video.
Problem
Creating photorealistic head avatars that are both novel-view renderable and controllable via expression, pose, and viewpoint, which NeRF-based and deformation-field methods struggle to combine.
Approach
Attach one 3D Gaussian to each FLAME triangle in the triangle's local frame so Gaussians move with the parametric mesh; add binding inheritance so densified Gaussians stay rigged, and regularize Gaussian position and scale relative to the parent triangle to avoid animation artifacts; jointly fine-tune FLAME parameters with the splats.
Key results
- Novel-view synthesis on NeRSemble: PSNR 31.6, SSIM 0.938, LPIPS 0.065, beating INSTA (26.7 PSNR) and AvatarMAV (29.5 PSNR).
- Self-reenactment: LPIPS 0.076 vs 0.110 for INSTA, though PSNR 26.0 is below INSTA's 26.3 in the same setting.
- Ablation on subject 304 shows removing adaptive density control with binding inheritance drops novel-view PSNR from 28.8 to 26.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GaussianAvatars instead of the strongest listed baseline, AvatarMAV | Head avatar reconstruction · Novel-View Synthesis | PSNR: 29.5 -> 31.6 | Helped |
| Use GaussianAvatars instead of the strongest listed baseline, INSTA | Head avatar reconstruction · Self-Reenactment | LPIPS: 0.110 -> 0.076 | Helped |
| Add adaptive density control with binding inheritance (ablation) | GaussianAvatars · Novel-View Synthesis | LPIPS: 0.206 -> 0.098 | Helped |
| Add adaptive density control with binding inheritance (ablation) | GaussianAvatars · Self-Reenactment | LPIPS: 0.183 -> 0.101 | Helped |
| Add Gaussian scaling loss (ablation) | GaussianAvatars · Novel-View Synthesis | LPIPS: 0.114 -> 0.098 | Helped |
| Keep the scaling-loss tolerance εscaling (ablation) | GaussianAvatars · Novel-View Synthesis | PSNR: 25.0 -> 28.8 | Helped |
| Add Gaussian position loss (ablation) | GaussianAvatars · Novel-View Synthesis | PSNR: 29.7 -> 28.8 | Hurt |
| Add Gaussian position loss (ablation) | GaussianAvatars · Self-Reenactment | LPIPS: 0.101 -> 0.096 | Helped |
| Keep the position-loss threshold εposition (ablation) | GaussianAvatars · Novel-View Synthesis | PSNR: 28.7 -> 28.8 | No clear effect |
| Fine-tune FLAME parameters during Gaussian optimization (ablation) | GaussianAvatars · Novel-View Synthesis | PSNR: 26.1 -> 28.8 | Helped |
| Fine-tune FLAME parameters during Gaussian optimization (ablation) | GaussianAvatars · Self-Reenactment | LPIPS: 0.124 -> 0.101 | Helped |
Surprising findings
- Removing the position loss improves novel-view PSNR (29.7 vs 28.8) while harming animation quality, showing a tradeoff between fidelity and robustness.
Limitations
- Cannot relight the avatar because appearance is not decoupled from lighting.
- Hair and accessories not modeled by FLAME are not controllable.
- Self-reenactment PSNR is lower than INSTA because tracked meshes may not align perfectly with target images.
Experimental setup
Datasets: NeRSemble
Benchmarks: NeRSemble
Metrics: PSNR SSIM LPIPS
Baselines: AvatarMAV PointAvatar INSTA
Builds on: 3D Gaussian Splatting FLAME adaptive density control
Closest papers by meaning
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.92)
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.91)
- Generalizable and Animatable Gaussian Head Avatar (NeurIPS 2024 · similarity 0.91)
- 3D Gaussian Head Avatars with Expressive Dynamic Appearances by Compact Tensorial Representations (CVPR 2025 · similarity 0.91)
- GAvatar: Animatable 3D Gaussian Avatars with Implicit Mesh Learning (CVPR 2024 · similarity 0.90)
- GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians (CVPR 2024 · 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.