Research map / 3D vision / Gaussian splatting
3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting
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.
Creates animatable clothed human avatars from monocular videos by deforming canonical 3D Gaussians with a pose-conditioned non-rigid MLP, LBS rigid articulation, a shallow color MLP, and as-isometric-as-possible regularization on Gaussian positions and covariances.
Problem
Animatable clothed human avatar reconstruction from monocular video that is simultaneously fast to train, renders in real time, and models pose-dependent non-rigid cloth deformation while generalizing to unseen poses.
Approach
Applies 3D Gaussian Splatting to articulated human avatars by combining a hash-grid pose-conditioned non-rigid deformation network with forward LBS via a learned skinning field (no inverse skinning), replaces SH color with a tiny MLP using canonicalized view direction, and adds isometry regularization on Gaussian means and covariances.
Key results
- On ZJU-MoCap, 3DGS-Avatar achieves LPIPS 29.58 (ablation full model) and comparable PSNR/SSIM to ARAH, trains in about 30 GPU minutes versus 8 GPU days, and renders at 50 FPS.
- On PeopleSnapshot, it significantly outperforms InstantAvatar on PSNR and LPIPS (e.g., female-3-casual PSNR 30.57 vs 27.66) while being over 3x faster at inference.
- Ablation shows removing the color MLP raises LPIPS from 29.58 to 31.24, and removing isometric regularization raises it to 30.25.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 3DGS-Avatar instead of ARAH | Animatable avatar reconstruction · ZJU-MoCap | LPIPS: 26.60 -> 20.88; 31.40 -> 25.77; 40.43 -> 34.24; 35.28 -> 30.14; 42.30 -> 35.26; 40.76 -> 31.21 | Helped |
| Use 3DGS-Avatar instead of InstantAvatar | Novel-pose avatar synthesis · PeopleSnapshot | PSNR: 27.66 -> 30.57; 29.11 -> 33.16; 29.53 -> 34.28; 27.67 -> 30.22 | Helped |
| Use 3DGS-Avatar instead of InstantAvatar | Novel-pose avatar synthesis · PeopleSnapshot | LPIPS: 21.00 -> 20.86; 16.70 -> 15.74; 15.50 -> 14.92; 30.7 -> 23.05 | Mixed |
| Remove color MLP (ablation) | 3DGS-Avatar · ZJU-MoCap | LPIPS: 29.58 -> 31.24 | Hurt |
| Add covariance regularization (ablation) | 3DGS-Avatar · ZJU-MoCap | LPIPS: 29.84 -> 29.58 | Helped |
| Add position and covariance regularization (ablation) | 3DGS-Avatar · ZJU-MoCap | LPIPS: 30.25 -> 29.58 | Helped |
| Add pose correction (ablation) | 3DGS-Avatar · ZJU-MoCap | LPIPS: 29.87 -> 29.58 | Helped |
Surprising findings
- Rendering quality is better measured by LPIPS than PSNR/SSIM in the monocular setting, where PSNR/SSIM differences are small across methods.
Limitations
- Loss-function weights and detailed definitions are deferred to supplementary material
- Quantitative evaluation focuses on ZJU-MoCap with repetitive motion; novel pose synthesis is only shown qualitatively
- Instant-NVR is trained on a refined dataset so its metrics are not directly comparable
Experimental setup
Datasets: ZJU-MoCap PeopleSnapshot AMASS AIST++
Benchmarks: ZJU-MoCap PeopleSnapshot
Metrics: PSNR SSIM LPIPS FPS training time
Baselines: NeuralBody HumanNeRF MonoHuman ARAH Instant-NVR InstantAvatar
Builds on: 3D Gaussian Splatting HumanNeRF Instant-NGP as-isometric-as-possible deformation SMPL
Closest papers by meaning
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.94)
- ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering (CVPR 2024 · similarity 0.93)
- GAvatar: Animatable 3D Gaussian Avatars with Implicit Mesh Learning (CVPR 2024 · similarity 0.93)
- AniGS: Animatable Gaussian Avatar from a Single Image with Inconsistent Gaussian Reconstruction (CVPR 2025 · similarity 0.92)
- GASP: Gaussian Avatars with Synthetic Priors (CVPR 2025 · similarity 0.92)
- SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting (CVPR 2024 · similarity 0.91)
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.