Research map / 3D vision / Face and shape
SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic Priors
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 SinGS, which samples kinematic (pose-space) human image sequences with a video diffusion model (KHD) and distills them into an animatable 3D Gaussian avatar using semantic Laplacian regularization and geometry-preserving density control.
Problem
Single-image animatable 3D human avatar reconstruction suffers from incomplete observation (occluded views, no motion info) and inconsistent 3D priors from SDS-based diffusion, and existing methods are slow or static.
Approach
Replace camera-space view hallucination with pose-space sampling: a diffusion model conditioned on SMPL pose renderings generates a 3D-consistent pose sequence of the subject, and a 3DGS avatar is built on a triplane over a loop-subdivided template with semantic Laplacian regularization and mesh-based density control to tolerate imperfect priors.
Key results
- SinGS achieves PSNR 31.33, LPIPS 0.240, FVD 156.3 on UBC Fashion while running at 70 FPS with 3.8 GB GPU memory.
- KHD alone reaches PSNR 32.81 and FVD 134.4, outperforming Animate Anyone and Champ on the UBC Fashion test set, but at only 2 FPS.
- Semantic Laplacian regularization improves PSNR from 30.88 (KNN Laplacian) to 31.33, and GPDC improves PSNR over naive density control from 30.22 to 31.01.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Semantic Laplacian Regularization instead of KNN Laplacian Regularization (ablation) | SinGS · UBC fashion dataset | PSNR: 30.88 -> 31.33 | Helped |
| add Semantic Laplacian Regularization (ablation) | SinGS · UBC fashion dataset | PSNR: 30.34 -> 31.33 | Helped |
| add Semantic Laplacian Regularization (ablation) | SinGS · UBC fashion dataset | LPIPS: 0.272 -> 0.240 | Helped |
| add Semantic Laplacian Regularization (ablation) | SinGS · UBC fashion dataset | FVD: 181.3 -> 156.3 | Helped |
| use Geometry-Preserving Density Control instead of naive Density Control (ablation) | SinGS · UBC fashion dataset | PSNR: 30.22 -> 31.01 | Helped |
| use Geometry-Preserving Density Control instead of naive Density Control (ablation) | SinGS · UBC fashion dataset | LPIPS: 0.276 -> 0.249 | Helped |
| use Geometry-Preserving Density Control instead of naive Density Control (ablation) | SinGS · UBC fashion dataset | FVD: 190.2 -> 166.1 | Helped |
Surprising findings
- Increasing the number of sampled poses beyond about 100 yields diminishing returns and can harm performance because inconsistent generated frames add noise.
- SinGS lags behind the diffusion-based UniAnimate in PSNR, FVD and CLIP score, trading some quality for speed.
Limitations
- KHD inference is slow (about 2 FPS) and resource-heavy, so prior generation is the bottleneck.
- Generated priors retain some 3D inconsistencies that must be tolerated by the reconstruction stage.
- Evaluation is limited to the UBC Fashion dataset with 100 test videos plus a 50-user study.
Experimental setup
Datasets: RotationHuman UBC Fashion MVHumanNet TikTok
Benchmarks: UBC Fashion
Metrics: PSNR LPIPS FVD Contextual Distance CLIP score FPS GPU memory user study score
Baselines: DisCo Animate Anyone Champ UniAnimate SHERF SIFU GST TeCH SiTH HumanRef SHERT HAHA
Builds on: 3D Gaussian Splatting UniAnimate HUGS Score Distillation Sampling Linear Blend Skinning
Closest papers by meaning
- IDOL: Instant Photorealistic 3D Human Creation from a Single Image (CVPR 2025 · similarity 0.88)
- AniGS: Animatable Gaussian Avatar from a Single Image with Inconsistent Gaussian Reconstruction (CVPR 2025 · similarity 0.88)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.88)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.87)
- Human-3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models (NeurIPS 2024 · similarity 0.87)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.86)
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.