Research map / 3D vision / Gaussian splatting
ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering
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 ASH, an animatable Gaussian splatting method that attaches 3D Gaussians to a deformable template mesh and predicts their parameters in UV texture space with 2D convolutional decoders, enabling real-time photoreal human rendering from multi-view video.
Problem
Real-time, photorealistic, pose-controllable rendering of clothed human avatars, where naive 3D Gaussian parameter prediction from skeletal motion is too costly and hybrid NeRF-based methods are too slow.
Approach
Parameterize a fixed number of 3D Gaussians as texels on the UV map of a deformable template mesh, so the mapping from skeletal motion (via motion-aware normal and position textures) to Gaussian parameters becomes a 2D image-to-image translation solved by efficient 2D CNNs, with a two-stage warmup and final training supervised by multi-view videos.
Key results
- On novel view synthesis (tight outfits), ASH reaches 35.84 PSNR and 11.92 LPIPS, versus 31.21 PSNR for DDC, the best real-time competitor.
- On novel pose synthesis (loose outfits), ASH reaches 27.12 PSNR and 20.22 LPIPS, outperforming HDHumans (26.71 PSNR, 22.75 LPIPS) which is non-real-time.
- ASH renders at about 30 fps (29.64 fps reported) at 1K resolution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ASH instead of HDHumans | Animatable human rendering · DynaCap novel view, Tight Outfits | PSNR: 30.98 -> 35.84 | Helped |
| use ASH instead of HDHumans | Animatable human rendering · DynaCap novel view, Loose Outfits | LPIPS: 15.79 -> 8.30 | Helped |
| use ASH instead of the strongest baseline | Animatable human rendering · DynaCap novel pose, Tight Outfits | PSNR: 28.78 -> 28.90 | Helped |
| use ASH instead of HDHumans | Animatable human rendering · DynaCap novel pose, Loose Outfits | LPIPS: 22.75 -> 20.22 | Helped |
| add motion conditioning (ablation) | ASH · DynaCap novel pose, Loose Outfits | LPIPS: 32.35 -> 20.22 | Helped |
| add the motion-aware Gaussian offset (ablation) | ASH · DynaCap novel pose, Loose Outfits | LPIPS: 25.32 -> 20.22 | Helped |
| reduce texture resolution to 128 (ablation) | ASH · DynaCap novel pose, Loose Outfits | LPIPS: 20.22 -> 22.00 | Hurt |
| increase texture resolution to 512 (ablation) | ASH · DynaCap training pose, Loose Outfits | PSNR: 35.47 -> 35.28 | No clear effect |
Limitations
- The underlying deformable template mesh is not updated by the Gaussian splatting.
- Quality depends on the texture resolution: halving it to 128 causes blurry renderings and doubling it to 512 breaks real-time performance.
Experimental setup
Datasets: DynaCap newly recorded multi-view sequences
Benchmarks: DynaCap
Metrics: PSNR LPIPS
Baselines: DDC TAVA NA HDHumans
Builds on: 3D Gaussian splatting DDC HDHumans Dual Quaternion skinning
Closest papers by meaning
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.95)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.93)
- GASP: Gaussian Avatars with Synthetic Priors (CVPR 2025 · similarity 0.91)
- SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting (CVPR 2024 · similarity 0.90)
- HUGS: Human Gaussian Splats (CVPR 2024 · similarity 0.90)
- Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling (CVPR 2024 · similarity 0.90)
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.