Research map / 3D vision / Gaussian splatting
GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections
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 GPAvatar, a Gaussian splatting head avatar method that lifts Gaussians into an expression-conditioned high-dimensional space and uses a learned linear projection plus adaptive densification to get fast, high-fidelity monocular avatars.
Problem
Reconstructing high-fidelity animatable 3D head avatars from monocular video with high rendering speed and low memory, avoiding reliance on explicit priors or heavy neural implicit fields.
Approach
Extend each 3D Gaussian to an (n+3)-dimensional joint space with the FLAME expression code, then condition on the expression to obtain a 3D mean via a learnable linear map (3n parameters per Gaussian) while keeping covariance fixed; add a gradient-driven densification that also splits the linear map parameters by expression variance.
Key results
- On the INSTA dataset, GPAvatar reaches average PSNR 32.54 vs 31.33 for GaussianBlendShapes (Table 1 per-subject table, averaged).
- Renders at 450 FPS at 512x512 with 2.5G training memory and 1.5G runtime memory on an RTX 4090 (Table 3), versus 370 FPS for GaussianBlendShapes and 100 FPS for FlashAvatar.
- On the self-collected 1024x1024 dataset, average PSNR 29.92 vs 29.49 for GaussianBlendShapes and LPIPS 0.0707 vs 0.1119 (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GPAvatar instead of GaussianBlendShapes | Head avatar reconstruction · INSTA and GBS datasets (Avg.) | PSNR: 31.33 -> 32.54 | Helped |
| use GPAvatar instead of FlashAvatar | Head avatar reconstruction · INSTA and GBS datasets (Avg.) | LPIPS: 0.0871 -> 0.0551 | Helped |
| use GPAvatar instead of GaussianBlendShapes | Head avatar reconstruction · Authors' dataset (Avg.) | SSIM: 0.9369 -> 0.9505 | Helped |
| use GPAvatar instead of FlashAvatar | Head avatar reconstruction · Authors' dataset (Avg.) | LPIPS: 0.0785 -> 0.0707 | Helped |
| use GPAvatar instead of GaussianBlendShapes | Head avatar rendering · 512×512 rendering | rendering speed: 370fps -> 450fps | Helped |
| replace the linear transformation on µ with a 3-layer MLP (ablation) | GPAvatar · INSTA dataset | PSNR: 31.93 -> 31.74 | Hurt |
| use a 3-layer MLP to predict covariance changes (ablation) | GPAvatar · INSTA dataset | PSNR: 31.93 -> 32.03 | Mixed |
| use a 3-layer MLP to predict covariance changes (ablation) | GPAvatar · INSTA dataset | rendering speed: 450fps -> 120fps | Hurt |
| use 3-layer MLPs for position and covariance changes (ablation) | GPAvatar · INSTA dataset | PSNR: 31.93 -> 31.82 | Hurt |
| use 3-layer MLPs for position and covariance changes (ablation) | GPAvatar · INSTA dataset | rendering speed: 450fps -> 140fps | Hurt |
| remove sigmoid activation from the linear transformation (ablation) | GPAvatar · INSTA dataset | reconstruction quality | Hurt |
| remove L2 regularization from the linear transformation parameters (ablation) | GPAvatar · INSTA dataset | reconstruction quality | Hurt |
Surprising findings
- More complex MLP replacements for the linear transformation (on position, covariance, or both) did not improve quality and increased cost, supporting the simple linear design.
- The method lacks parametric head-model geometry priors and shows poor robustness to large pose and expression extrapolation.
Limitations
- No geometry prior from parametric head models, causing poor robustness to expression and pose extrapolation.
- Quantitative gains over prior methods are modest.
- Self-collected dataset is not released.
Experimental setup
Datasets: INSTA dataset GBS dataset self-collected 7-actor dataset
Benchmarks: INSTA dataset GBS dataset self-collected 7-actor dataset
Metrics: PSNR SSIM LPIPS rendering speed (FPS) training memory runtime memory
Baselines: PointAvatar INSTA FlashAvatar GaussianBlendShapes NeRFBlendShape
Builds on: 3D Gaussian Splatting 4D Gaussian Splatting Metrical Photometric Tracker FLAME
Closest papers by meaning
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.94)
- GraphAvatar: Compact Head Avatars with GNN-Generated 3D Gaussians (AAAI 2025 · similarity 0.92)
- GPAvatar: Generalizable and Precise Head Avatar from Image(s) (ICLR 2024 · similarity 0.92)
- FlashAvatar: High-fidelity Head Avatar with Efficient Gaussian Embedding (CVPR 2024 · similarity 0.92)
- HRAvatar: High-Quality and Relightable Gaussian Head Avatar (CVPR 2025 · similarity 0.91)
- Generalizable and Animatable Gaussian Head Avatar (NeurIPS 2024 · similarity 0.91)
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.