Research map / 3D vision / Gaussian splatting
GraphAvatar: Compact Head Avatars with GNN-Generated 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.
Proposes GraphAvatar, which uses Graph U-Nets to generate 3D Gaussian attributes from tracked FLAME meshes so only the small GNN weights (~10MB) are stored, plus graph-guided pose/expression refinement and a depth-modulated 3D-aware enhancer.
Problem
Head avatar rendering with 3D Gaussian Splatting requires large, fluctuating storage because Gaussians are attached per mesh triangle and densified, and results depend heavily on accurate face tracking.
Approach
Replace stored per-Gaussian parameters with geometric and appearance Graph U-Nets that generate Gaussian anchors from the tracked mesh, add neural Gaussians via view-dependent MLPs, jointly refine FLAME expression and camera pose with a graph-guided cross-attention module, and apply a depth-conditioned (FiLM-like) enhancer as post-processing.
Key results
- On INSTA, GraphAvatar reaches PSNR 30.52, SSIM 0.966, LPIPS 0.036 and 10.8 MB model size, versus 29.71 PSNR and 43.1 MB for the next best Gaussian Head Avatar.
- On NBS, GraphAvatar reaches PSNR 27.32 and LPIPS 0.061, compared with 26.38 PSNR for Gaussian Head Avatar.
- Rendering takes 0.029 s per frame, slower than GaussianAvatars (0.006 s) and FlashAvatar (0.007 s), but storage is the smallest among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace directly stored 3D Gaussians with GNN-generated Gaussians | GraphAvatar · INSTA | Size (MB): 13.3 -> 10.8 | Helped |
| use GraphAvatar instead of Gaussian Head Avatar | Head avatar animation · INSTA | PSNR: 29.71 -> 30.52 | Helped |
| use GraphAvatar instead of Gaussian Head Avatar | Head avatar animation · NBS | PSNR: 26.38 -> 27.32 | Helped |
| add the warm-up stage using pseudo Gaussians (ablation) | GraphAvatar · INSTA | PSNR: 8.97 -> 30.52 | Helped |
| add neural Gaussians to represent elements missing from the FLAME template (ablation) | GraphAvatar · INSTA | PSNR: 29.61 -> 30.52 | Helped |
| add graph-guided optimization to refine tracked parameters (ablation) | GraphAvatar · INSTA | PSNR: 30.41 -> 30.52 | Helped |
| add the 3D-aware enhancer (ablation) | GraphAvatar · INSTA | SSIM: 0.959 -> 0.966 | Helped |
| add the 3D-aware enhancer (ablation) | GraphAvatar · INSTA | LPIPS: 0.042 -> 0.036 | Helped |
Surprising findings
- Without the warm-up stage, training is unstable and fails to converge (PSNR 8.97 on INSTA), showing graph networks are hard to optimize from scratch.
Limitations
- Rendering time (0.029 s) is slower than some 3DGS baselines such as GaussianAvatars and FlashAvatar.
- Training is unstable without a warm-up stage that relies on pseudo Gaussians from vanilla 3DGS.
- FLAME-based mesh cannot model teeth, hair, and accessories well, requiring neural Gaussians to cover them.
Experimental setup
Datasets: NeRFBlendShape dataset INSTA dataset
Benchmarks: INSTA dataset NeRFBlendShape (NBS) dataset
Metrics: L2 PSNR SSIM LPIPS rendering time model size (MB)
Baselines: NHA IMAvatar INSTA FlashAvatar Gaussian Head Avatar GaussianAvatars NeRFBlendShape
Builds on: 3D Gaussian Splatting Scaffold-GS Graph U-Net FLAME Chebyshev graph convolution
Closest papers by meaning
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.92)
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.90)
- HRAvatar: High-Quality and Relightable Gaussian Head Avatar (CVPR 2025 · similarity 0.89)
- GPAvatar: Generalizable and Precise Head Avatar from Image(s) (ICLR 2024 · similarity 0.89)
- FlashAvatar: High-fidelity Head Avatar with Efficient Gaussian Embedding (CVPR 2024 · similarity 0.89)
- RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars (CVPR 2025 · similarity 0.89)
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.