Research map / 3D vision / Face and shape
GART: Gaussian Articulated Template Models
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.
Introduces GART, an explicit mixture-of-3D-Gaussians representation rigged to a template model (SMPL/SMAL) with learnable forward skinning and latent bones, reconstructed from monocular videos.
Problem
Reconstructing and rendering non-rigid articulated subjects (humans, animals) with detailed appearance and geometry from monocular video, where implicit neural fields are slow to render and explicit representations have low quality.
Approach
Approximate the canonical radiance field with a Gaussian mixture initialized on a template mesh, deform it by learnable forward skinning plus unobserved latent bones with learnable skinning weights, and regularize with voxel-distilled skinning and KNN smoothness, rendered via 3D Gaussian splatting.
Key results
- On ZJU-MoCap, GART reaches 32.22 PSNR / 0.977 SSIM in ~2.5 min training and 31.76 PSNR in ~30 s, outperforming Instant-NVR (31.01 PSNR).
- On UBC-Fashion, GART achieves 25.65 PSNR vs. 19.97 for InstantAvatar (training pose), and ablations show latent bones and learnable skinning both help.
- GART renders at more than 150 FPS at 540x540 on a single RTX-3080-Laptop GPU, versus about 15 FPS for Instant-Avatar.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GART instead of Instant-NVR | Monocular human rendering with SMPL · ZJU-MoCap | PSNR: 31.01 -> 32.22 | Helped |
| Reduce GART training time from ∼2.5min to ∼30s | GART · ZJU-MoCap | PSNR: 32.22 -> 31.76 | Hurt |
| Use GART instead of InstantAvatar | Monocular human rendering with SMPL · People-Snapshot, male-3-casual | PSNR: 29.65 -> 30.40 | Helped |
| Use a per-frame table instead of an MLP for latent bone transformations (ablation) | GART · UBC-Fashion | PSNR: 25.65 -> 25.96 | Helped |
| Add learnable skinning (ablation) | GART · UBC-Fashion | PSNR: 23.76 -> 25.65 | Helped |
| Add latent bones (ablation) | GART · UBC-Fashion | PSNR: 25.00 -> 25.65 | Helped |
| Add latent bones (ablation) | GART · UBC-Fashion | LPIPS: 82.03 -> 81.88 | Helped |
| Add voxel-distilled skinning weights (ablation) | GART · In-the-wild video, side view | visual artifacts | Helped |
| Add KNN regularization (ablation) | GART · In-the-wild video, back view | visual artifacts | Helped |
Limitations
- Relies on a template pose estimator, which may not exist for general animal species.
- Fits a single monocular video per subject; does not learn category-level priors from video collections.
Experimental setup
Datasets: ZJU-MoCap People-Snapshot UBC-Fashion National Dog Show videos Adobe Stock Videos
Benchmarks: ZJU-MoCap People-Snapshot UBC-Fashion ITW-Dogs
Metrics: PSNR SSIM LPIPS FPS training time
Baselines: Instant-Avatar Instant-NVR HumanNeRF Neural Body Anim-NeRF NeuralBody/AS/NB/AN/NHP/PixelNeRF comparisons on ZJU-MoCap
Builds on: 3D Gaussian Splatting SMPL D-SMAL Fast-SNARF
Closest papers by meaning
- GauHuman: Articulated Gaussian Splatting from Monocular Human Videos (CVPR 2024 · similarity 0.88)
- RigGS: Rigging of 3D Gaussians for Modeling Articulated Objects in Videos (CVPR 2025 · similarity 0.87)
- REArtGS: Reconstructing and Generating Articulated Objects via 3D Gaussian Splatting with Geometric and Motion Constraints (NeurIPS 2025 · similarity 0.85)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.85)
- TAGA: Self-supervised Learning for Template-free Animatable Gaussian Articulated Model (CVPR 2025 · similarity 0.85)
- Building Interactable Replicas of Complex Articulated Objects via Gaussian Splatting (ICLR 2025 · similarity 0.85)
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.