Research map / 3D vision / Face and shape
High-Fidelity Mobile Avatars with Pruned Local Blendshapes
Youyi Zhan, He Wang, Tianjia Shao, Kun Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add local per-part pose features and blendshapes instead of one global pose feature (length 16) (ablation) | Ours · AvatarRex (training pose, novel view) | PSNR: 27.645 -> 29.405 | Improved |
| add local per-part features instead of a global feature of length 64 (ablation) | Ours · AvatarRex (training pose, novel view) | PSNR: 27.913 -> 29.405 | Improved |
| add local per-part features (LPIPS vs global 16) (ablation) | Ours · AvatarRex (training pose, novel view) | LPIPS: 0.1276 -> 0.1074 | Improved |
| add blendshape pruning plus fine-tuning (vs no pruning) (ablation) | Ours · AvatarRex (training pose, novel view) | PSNR: 29.342 -> 29.405 | No clear effect |
| add blendshape pruning (model size) (ablation) | Ours · RTX 3050 2K rendering | Model size (MB): 72.4 -> 19.5 | Improved |
| add blendshape pruning (GPU memory) (ablation) | Ours · RTX 3050 2K rendering | GPU memory (MB): 155 -> 105 | Improved |
| add blendshape pruning (speed) (ablation) | Ours · RTX 3050 2K rendering | FPS on RTX 3050: 191 -> 312 | Improved |
| add blendshape pruning (blendshape combination time) (ablation) | Ours · RTX 3050 2K rendering | blendshape combination time per frame (ms): 2.26 ms -> 0.52 ms | Improved |
| group Gaussians by body-part locality (local PCA) rather than global PCA, 16 components (ablation) | PCA of color correctives · PCA of AnimatableGS color correctives | L1 reconstruction error: 0.0159 -> 0.0042 | Improved |
| use global PCA with 64 components instead of 16 | PCA of color correctives · PCA of AnimatableGS color correctives | L1 reconstruction error: 0.0159 -> 0.0092 | Improved |
| group Gaussians by position instead of randomly for local PCA (ablation) | PCA of color correctives · PCA of AnimatableGS color correctives | L1 reconstruction error: 0.0158 -> 0.0042 | Improved |
| add dynamic Gaussian prediction (vs static Gaussians only) (ablation) | Ours · RTX 3050 2K rendering | FPS on RTX 3050: 410 -> 312 | Hurt |
Closest papers by meaning
- Real-time High-fidelity Gaussian Human Avatars with Position-based Interpolation of Spatially Distributed MLPs (CVPR 2025 · similarity 0.91)
- RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars (CVPR 2025 · similarity 0.90)
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.88)
- GAF: Gaussian Avatar Reconstruction from Monocular Videos via Multi-view Diffusion (CVPR 2025 · similarity 0.88)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.87)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.87)
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.
