Research map / 3D vision / Face and shape
GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians
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 GaussianAvatar, animatable 3D Gaussians with pose-dependent appearance network and jointly optimized motion to reconstruct human avatars from a single monocular video.
Problem
Creating realistic animatable human avatars from a single monocular video, where inaccurate motion estimation and limited pose coverage cause artifacts and poor generalization.
Approach
Represent humans with SMPL-anchored animatable 3D Gaussians repositioned by forward skinning, predict pose-dependent Gaussian properties from a UV positional map via a pose encoder plus an optimizable feature tensor, and jointly optimize SMPL pose and translation with the appearance through the differentiable forward-skinning path.
Key results
- On DynVideo, Baseline + Opt. + Dyn. reaches 28.58 PSNR / 0.9616 SSIM / 0.0169 LPIPS versus 24.39 PSNR for HumanNeRF and 20.31 for InstantAvatar.
- On NeuMan, the full method attains 29.94 PSNR and 0.0124 LPIPS, above HumanNeRF (27.06) and InstantAvatar (28.47).
- Joint motion optimization gives the largest gain on People-Snapshot and NeuMan, e.g. NeuMan PSNR rising from 27.06 (baseline) to 29.93.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add joint motion optimization to the baseline (ablation) | GaussianAvatar · People-Snapshot | PSNR: 27.71 -> 30.98; 25.09 -> 28.77; 24.80 -> 29.59; 26.16 -> 30.83 | Helped |
| Add joint motion optimization to the baseline (ablation) | GaussianAvatar · NeuMan | PSNR: 27.06 -> 29.93 | Helped |
| Add joint motion optimization to the baseline (ablation) | GaussianAvatar · DynVideo | PSNR: 23.55 -> 26.16 | Helped |
| Add dynamic appearance modeling to motion-optimized GaussianAvatar (ablation) | GaussianAvatar · People-Snapshot | PSNR: 30.98 -> 30.98; 28.77 -> 28.78; 29.59 -> 29.55; 30.83 -> 30.84 | Mixed |
| Add dynamic appearance modeling to motion-optimized GaussianAvatar (ablation) | GaussianAvatar · NeuMan | PSNR: 29.93 -> 29.94 | No clear effect |
| Add dynamic appearance modeling to motion-optimized GaussianAvatar (ablation) | GaussianAvatar · DynVideo | PSNR: 26.16 -> 28.58 | Helped |
Surprising findings
- On People-Snapshot, the dynamic appearance modeling yields essentially no gain over motion optimization alone (differences within about 0.01 dB), attributed to limited pose variation.
Limitations
- Artifacts from inaccurate foreground segmentation in videos.
- Difficulty modeling loose outfits such as dresses.
Experimental setup
Datasets: People-Snapshot NeuMan DynVideo
Benchmarks: People-Snapshot NeuMan DynVideo
Metrics: PSNR SSIM LPIPS
Baselines: HumanNeRF InstantAvatar
Builds on: 3D Gaussian Splatting SMPL HumanNeRF InstantAvatar
Closest papers by meaning
- Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling (CVPR 2024 · similarity 0.91)
- GAvatar: Animatable 3D Gaussian Avatars with Implicit Mesh Learning (CVPR 2024 · similarity 0.90)
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · similarity 0.90)
- GoMAvatar: Efficient Animatable Human Modeling from Monocular Video Using Gaussians-on-Mesh (CVPR 2024 · similarity 0.90)
- MPMAvatar: Learning 3D Gaussian Avatars with Accurate and Robust Physics-Based Dynamics (NeurIPS 2025 · similarity 0.90)
- Real-time High-fidelity Gaussian Human Avatars with Position-based Interpolation of Spatially Distributed MLPs (CVPR 2025 · similarity 0.89)
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.