Research map / 3D vision / Gaussian splatting
3D Gaussian Head Avatars with Expressive Dynamic Appearances by Compact Tensorial Representations
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 a compact 3DGS head avatar that stores neutral appearance in canonical tri-planes and expression-dependent opacity offsets in 1D feature lines bound to FLAME blendshapes, with adaptive truncated opacity penalty and class-balanced sampling.
Problem
Capturing dynamic facial textures in 3DMM-driven Gaussian head avatars while keeping real-time rendering speed and low storage, and generalizing to unseen expressions.
Approach
Encode neutral-face texture in tri-planes replacing spherical harmonics, and represent per-blendshape dynamic texture as 1D feature lines interpolated by blendshape coefficients and decoded by an MLP into opacity offsets; add an adaptive truncated opacity penalty on static triangles and class-balanced resampling over expression clusters.
Key results
- Achieves PSNR 32.97 on novel view synthesis vs 31.47 for GA, with 10MB storage and 300 FPS on RTX 4090.
- Self-reenactment PSNR 28.07, above GA (27.27) and GBS (25.98), but SSIM and LPIPS are not best across all baselines.
- GBS requires 2GB storage and GHA runs at 20 FPS, while GA uses 21MB at 330 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use compact tensorial representations and dynamic feature lines | Ours · Novel View Synthesis | PSNR: 31.4702 -> 32.9688 | Helped |
| use compact tensorial representations and dynamic feature lines | Ours · Self-Reenactment | PSNR: 27.2678 -> 28.0688 | Helped |
| use compact tensorial representations and dynamic feature lines | Ours · Novel View Synthesis | LPIPS: 0.05144 -> 0.05940 | Hurt |
| use compact tensorial representations and dynamic feature lines | Ours · Self-Reenactment | LPIPS: 0.06668 -> 0.07724 | Hurt |
| use compact tensorial representations and dynamic feature lines | Ours · Novel View Synthesis | Storage: 21M -> 10M | Helped |
| add feature-line blendshapes (ablation) | Triplane-only method · Novel View Synthesis, subject #306 | PSNR: 33.4206 -> 33.4337 | Helped |
| add feature-line blendshapes (ablation) | Triplane-only method · Self-Reenactment, subject #306 | PSNR: 31.2487 -> 31.3246 | Helped |
| add adaptive opacity offset penalty (ablation) | Triplane + feature-line method · Novel View Synthesis, subject #306 | PSNR: 33.4337 -> 35.3876 | Helped |
| add adaptive opacity offset penalty (ablation) | Triplane + feature-line method · Self-Reenactment, subject #306 | PSNR: 31.3246 -> 31.4366 | Helped |
| add class-balanced resampling (ablation) | Triplane + feature-line method · Self-Reenactment, subject #306 | PSNR: 31.3246 -> 31.4288 | Helped |
| add class-balanced resampling (ablation) | Triplane + feature-line method · Novel View Synthesis, subject #306 | PSNR: 33.4337 -> 35.2652 | Mixed |
Surprising findings
- The triplane alone improves unseen-expression metrics but worsens LPIPS, indicating blurred high-frequency detail.
Limitations
- Relies on tracked FLAME mesh, so cannot handle complex hairstyles or mouth-interior topology changes.
- Does not decouple material and lighting, so relighting is not feasible.
- Evaluated only on nine Nersemble subjects.
Experimental setup
Datasets: Nersemble
Benchmarks: Nersemble
Metrics: PSNR SSIM LPIPS storage FPS
Baselines: Gaussian Avatars (GA) Gaussian Head Avatar (GHA) Gaussian Blendshapes (GBS)
Builds on: Gaussian Avatars 3D Gaussian Splatting FLAME
Closest papers by meaning
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.92)
- Gaussian Eigen Models for Human Heads (CVPR 2025 · similarity 0.91)
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · similarity 0.91)
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.90)
- MeGA: Hybrid Mesh-Gaussian Head Avatar for High-Fidelity Rendering and Head Editing (CVPR 2025 · similarity 0.90)
- Generalizable and Animatable Gaussian Head Avatar (NeurIPS 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.