Research map / 3D vision / Gaussian splatting
Human Gaussian Splatting: Real-time Rendering of Animatable Avatars
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 HuGS, an animatable human avatar built on 3D Gaussian Splatting that deforms canonical Gaussians with forward LBS plus a pose-dependent MLP refinement, enabling real-time novel-pose rendering from multi-view videos.
Problem
Real-time, photorealistic rendering of animatable human body avatars learned from multi-view video, with good generalization to novel body poses; neural NeRF/SDF avatars are slow to render and generalize poorly to unseen poses.
Approach
Represent the canonical body as 3D Gaussians with learnable per-Gaussian skinning weights and latent codes, deform them with direct forward linear blend skinning (avoiding backward skinning), then refine positions, rotations and colors with a shallow pose-dependent MLP plus regularizers that keep the MLP outputs small.
Key results
- On THuman4 novel pose synthesis, HuGS reaches 32.49 PSNR vs 30.97 for PoseVocab (about 1.5 dB gain) and renders 512x512 in ~12 ms (~80 fps) on a Tesla V100.
- On ZJU-MoCap novel pose synthesis, HuGS achieves the highest PSNR (23.69) among compared methods, while second-best on novel view PSNR (26.58 vs SLRF 28.32).
- Learned per-Gaussian skinning weights improve PSNR by about 1 dB over SMPL-imported weights (31.99 vs 30.97) on THuman4 novel views.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HuGS instead of PoseVocab | THuman4 subject00 · novel poses | PSNR: 30.97 -> 32.49 | Helped |
| use HuGS instead of DVA | DNA-Rendering 0165 · novel views | PSNR: 29.8 -> 31.5 | Helped |
| use HuGS instead of DVA | DNA-Rendering 0166 · novel poses | PSNR: 25.4 -> 25.7 | Helped |
| use HuGS instead of DVA | DNA-Rendering 0206 · novel poses | PSNR: 23.1 -> 23.2 | No clear effect |
| use HuGS instead of SLRF | ZJU-MoCap · novel poses | PSNR: 23.61 -> 23.69 | Helped |
| add learned per-Gaussian skinning weights instead of using closest-vertex SMPL weights (ablation) | HuGS with LBS only · THuman4 novel views | PSNR: 30.97 -> 31.99 | Helped |
| use per-Gaussian latent code instead of Fourier-encoded canonical position as MLP input (ablation) | HuGS · THuman4 novel pose synthesis | PSNR: 32.20 -> 32.49 | Helped |
| add the MLP shading output (ablation) | HuGS · THuman4 sequence00 novel pose synthesis | PSNR: 29.72dB -> 32.49 | Helped |
Surprising findings
- Using only LBS without the MLP is already a strong baseline quantitatively competitive with state-of-the-art methods.
Limitations
- Reconstruction degrades with sparse camera setups, causing overfitting and poor novel pose synthesis.
- The MLP fits observed garment deformations but does not extrapolate novel deformations to new poses.
- Each Gaussian is optimized and deformed independently, ignoring neighborhood structure between primitives.
Experimental setup
Datasets: THuman4 DNA-Rendering ZJU-MoCap
Benchmarks: THuman4 DNA-Rendering ZJU-MoCap
Metrics: PSNR SSIM LPIPS FID render time training time
Baselines: PoseVocab SLRF TAVA Ani-NeRF ARAH DVA Neural Body
Builds on: 3D Gaussian Splatting SMPL SCANimate SMPL-X
Closest papers by meaning
- HUGS: Human Gaussian Splats (CVPR 2024 · similarity 0.95)
- ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering (CVPR 2024 · similarity 0.95)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.94)
- SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting (CVPR 2024 · similarity 0.93)
- GASP: Gaussian Avatars with Synthetic Priors (CVPR 2025 · similarity 0.92)
- Gaussian Head & Shoulders: High Fidelity Neural Upper Body Avatars with Anchor Gaussian Guided Texture Warping (ICLR 2025 · similarity 0.92)
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.