Research map / 3D vision / Face and shape
Generalizable and Animatable Gaussian Head Avatar
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 GAGAvatar, a generalizable one-shot 3D Gaussian head avatar that predicts dual-lifted Gaussians and 3DMM-based expression Gaussians from a single image in one forward pass, enabling real-time reenactment.
Problem
One-shot animatable head avatar reconstruction from a single image with real-time reenactment, since prior NeRF-based methods are slow and prior 3DGS methods need per-identity training.
Approach
A dual-lifting method predicts forward and backward lifting distances of each pixel of a DINOv2 feature plane to form 3D Gaussians in one pass, constrained by 3DMM (FLAME) vertex priors; expression Gaussians are attached to learnable FLAME vertex features concatenated with global identity features, and a StyleUNet neural renderer refines coarse splats.
Key results
- On VFHQ self-reenactment, reaches PSNR 21.83, SSIM 0.818, LPIPS 0.122, outperforming the best baseline Portrait4D-v2 (21.34, 0.791, 0.144) on these metrics.
- On HDTF self-reenactment, achieves PSNR 23.13 and LPIPS 0.103, versus GPAvatar's 23.06 and 0.104 for the second-best PSNR.
- Runs at 67.12 FPS on an A100 GPU, versus at most 18.09 FPS for the fastest baseline (CVTHead).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GAGAvatar instead of Portrait4D-v2 | One-shot head avatar reenactment · VFHQ self reenactment | PSNR: 21.34 -> 21.83 | Helped |
| Use GAGAvatar instead of Portrait4D-v2 | One-shot head avatar reenactment · VFHQ self reenactment | LPIPS: 0.144 -> 0.122 | Helped |
| Use GAGAvatar instead of Real3DPortrait | One-shot head avatar reenactment · VFHQ cross reenactment | CSIM: 0.663 -> 0.633 | Hurt |
| Use GAGAvatar instead of GPAvatar | One-shot head avatar reenactment · HDTF self reenactment | PSNR: 23.06 -> 23.13 | Helped |
| Use GAGAvatar instead of Portrait4D-v2 | One-shot head avatar reenactment · HDTF self reenactment | SSIM: 0.860 -> 0.863 | Helped |
| Use GAGAvatar instead of Real3DPortrait | One-shot head avatar reenactment · HDTF cross reenactment | CSIM: 0.903 -> 0.851 | Hurt |
| Use GAGAvatar for reenactment inference | Avatar reenactment · A100 inference | FPS: 67.12 FPS | Helped |
| Add dual-lifting instead of one-plane lifting (ablation) | GAGAvatar · VFHQ self reenactment | PSNR: 21.34 -> 21.83 | Helped |
| Add global identity features to the expression branch (ablation) | GAGAvatar · VFHQ self reenactment | PSNR: 21.13 -> 21.83 | Helped |
| Add global identity features to the expression branch (ablation) | GAGAvatar · VFHQ cross reenactment | CSIM: 0.537 -> 0.633 | Helped |
| Add the neural renderer (ablation) | GAGAvatar · VFHQ self reenactment | PSNR: 20.34 -> 21.83 | Helped |
| Add lifting distance loss (ablation) | GAGAvatar · VFHQ self reenactment | PSNR: 21.64 -> 21.83 | Helped |
Surprising findings
- Cross-identity CSIM is slightly worse than some baselines (e.g., Real3DPortrait 0.663 vs. ours 0.633 on VFHQ), which the authors attribute to 3DMM identity/expression entanglement.
- Neural renderer does not upsample resolution; it mainly resolves conflicts between static dual-lifting and dynamic expression Gaussians.
Limitations
- Unseen regions (e.g., other half of face from side view, open mouth) lack detail and tend toward average expectations.
- 3DMM-based expression branch cannot control hair, tongue, or asymmetric eye states.
- Cross-identity identity consistency is slightly worse than some baselines due to 3DMM identity/expression entanglement.
Experimental setup
Datasets: VFHQ HDTF
Benchmarks: VFHQ HDTF
Metrics: PSNR SSIM LPIPS CSIM AED APD AKD FPS
Baselines: StyleHeat ROME OTAvatar HideNeRF GOHA CVTHead GPAvatar Real3DPortrait Portrait4D Portrait4D-v2
Builds on: 3D Gaussian splatting DINOv2 FLAME StyleUNet
Closest papers by meaning
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.91)
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.91)
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · similarity 0.91)
- RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars (CVPR 2025 · similarity 0.91)
- Gaussian Eigen Models for Human Heads (CVPR 2025 · similarity 0.91)
- 3D Gaussian Head Avatars with Expressive Dynamic Appearances by Compact Tensorial Representations (CVPR 2025 · 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.