Research map / 3D vision / Face and shape
GPAvatar: Generalizable and Precise Head Avatar from Image(s)
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 one-shot/few-shot generalizable 3D head avatar framework that uses a FLAME point-cloud-driven expression field and multi tri-planes attention fusion to reconstruct and reenact heads in a single forward pass.
Problem
Reconstructing animatable 3D head avatars from one or a few unseen portrait images with precise expression control, multi-view consistency, and generalization to new identities, without test-time optimization.
Approach
A dynamic point-based expression field attaches learnable features to FLAME vertices and queries neighbors globally with relative position encoding, so expressions move the feature field directly; a Multi Tri-planes Attention module fuses canonical tri-planes from an arbitrary number of input images.
Key results
- On VFHQ self-reenactment, one-input GPAvatar reaches PSNR 22.08 and LPIPS 0.177 vs. 20.07 and 0.204 for the best baseline HideNeRF.
- On HDTF self-reenactment, one-input GPAvatar reaches PSNR 24.21 and AKD 3.06 vs. 21.38 and 4.33 for HideNeRF.
- Two-input MTA improves VFHQ PSNR to 22.86 and Four-in to 23.49, and training takes about 50 GPU hours on 2 A100s with 15 FPS inference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GPAvatar One-in instead of the strongest listed baseline | GPAvatar · VFHQ self-reenactment | PSNR: 20.07 -> 22.08 | Helped |
| use GPAvatar One-in instead of the strongest listed baseline | GPAvatar · VFHQ self-reenactment | LPIPS: 0.204 -> 0.177 | Helped |
| use GPAvatar One-in instead of the strongest listed baseline | GPAvatar · VFHQ self-reenactment | AED: 0.497 -> 0.434 | Helped |
| use GPAvatar One-in instead of the strongest listed baseline | GPAvatar · VFHQ self-reenactment | AKD: 4.53 -> 3.53 | Helped |
| use GPAvatar One-in instead of HideNeRF | GPAvatar · HDTF self-reenactment | PSNR: 21.91 -> 24.21 | Helped |
| use GPAvatar One-in instead of HideNeRF | GPAvatar · HDTF self-reenactment | LPIPS: 0.147 -> 0.131 | Helped |
| use GPAvatar One-in instead of HideNeRF | GPAvatar · HDTF self-reenactment | AED: 0.499 -> 0.427 | Helped |
| use GPAvatar One-in instead of StyleHeat | GPAvatar · HDTF self-reenactment | AKD: 3.69 -> 3.06 | Helped |
| add the point-based expression field (ablation) | GPAvatar · VFHQ self-reenactment | AED: 0.576 -> 0.434 | Helped |
| use global rather than local point sampling (ablation) | GPAvatar PEF · VFHQ self-reenactment | AED: 0.518 -> 0.434 | Helped |
| reduce the point cloud to 2,000 points (ablation) | GPAvatar PEF · VFHQ self-reenactment | AED: 0.434 -> 0.497 | Hurt |
| reduce queried neighbors from 8 to 4 (ablation) | GPAvatar PEF · VFHQ self-reenactment | AED: 0.434 -> 0.514 | Hurt |
Surprising findings
- Cross-identity CSIM on VFHQ is slightly lower for GPAvatar (0.558) than HideNeRF (0.558 tie-level) and on HDTF HideNeRF has higher cross-identity CSIM (0.803 vs 0.790), indicating identity transfer is not uniformly best.
Limitations
- Shoulders and body below the neck are not controllable; shoulder position stays close to the input image.
- Hair, tongue and other non-FLAME regions cannot be explicitly controlled.
- Inference runs at about 15 FPS, below the real-time target of 30 FPS.
Experimental setup
Datasets: VFHQ HDTF
Benchmarks: VFHQ HDTF
Metrics: PSNR SSIM LPIPS CSIM L1 AED APD AKD
Baselines: ROME StyleHeat OTAvatar Next3D HideNeRF
Builds on: GFPGAN Point-NeRF NeRFace FLAME StyleGAN
Closest papers by meaning
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.92)
- GraphAvatar: Compact Head Avatars with GNN-Generated 3D Gaussians (AAAI 2025 · similarity 0.89)
- HRAvatar: High-Quality and Relightable Gaussian Head Avatar (CVPR 2025 · similarity 0.89)
- High-Fidelity 3D Head Avatars Reconstruction through Spatially-Varying Expression Conditioned Neural Radiance Field (AAAI 2024 · similarity 0.88)
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.88)
- Generalizable and Animatable Gaussian Head Avatar (NeurIPS 2024 · similarity 0.88)
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.