Research map / 3D vision / Face and shape
Gaussian Shell Maps for Efficient 3D Human Generation
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 Gaussian Shell Maps, a 3D GAN that generates articulable human bodies by sampling 3D Gaussians on SMPL-derived shells whose attributes come from CNN-generated texture maps, trained only on single-view images.
Problem
3D GAN generators for articulable digital humans rely on slow volume rendering or rasterized meshes with limited expressivity, and 2D upsamplers cause multi-view inconsistency; combining CNN generators with 3D Gaussian primitives in a generative setting was unexplored.
Approach
Anchor 3D Gaussians at fixed barycentric locations on multiple inflated/deflated SMPL shells and predict their color, opacity, scale and rotation from StyleGAN2-generated texture maps, so Gaussians are rendered by splatting and deformed with the SMPL skeleton; part-specific (face, hands, feet) and mask discriminators supervise training.
Key results
- Achieves FID 15.78 on DeepFashion and 13.30 on SHHQ at 512x512, comparable to SOTA (EVA3D 15.91 / 11.99).
- PCK of 99.48 (DeepFashion) and 99.27 (SHHQ), higher than all listed baselines.
- Inference at 28 ms/img on A6000 at 512x512, the fastest among compared methods; rendering reported at 125 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase the shell count from 1 to 2 (ablation) | Gaussian Shell Maps · SHHQ, 128² resolution | FID: 40.3 -> 25.31 | Helped |
| increase the shell count from 2 to 8 (ablation) | Gaussian Shell Maps · SHHQ, 128² resolution | FID: 25.31 -> 18.70 | Helped |
| increase the shell count from 8 to 10 (ablation) | Gaussian Shell Maps · SHHQ, 128² resolution | FID: 18.70 -> 21.57 | Hurt |
| increase the shell count from 8 to 15 (ablation) | Gaussian Shell Maps · SHHQ, 128² resolution | FID: 18.70 -> 21.10 | Hurt |
| add mask and part discriminators (ablation) | Gaussian Shell Maps · SHHQ, 128² resolution | FID: 20.63 -> 18.7 | Helped |
| sample Gaussians on shell meshes rather than in tetrahedra or uniformly inside a bounding box (ablation) | Gaussian Shell Maps · SHHQ | visual quality and convergence speed | Helped |
| use Gaussian Shell Maps instead of AG3D | DeepFashion · DeepFashion | FID: 10.93 -> 15.78 | Hurt |
| use Gaussian Shell Maps instead of StylePeople | DeepFashion · DeepFashion | PCK: 98.31 -> 99.48 | Helped |
| use Gaussian Shell Maps instead of EVA3D | SHHQ · SHHQ | FID: 11.99 -> 13.30 | Hurt |
| use Gaussian Shell Maps instead of GNARF | SHHQ · SHHQ | PCK: 98.96 -> 99.27 | Helped |
Surprising findings
- AG3D achieves the lowest FID on DeepFashion (10.93) yet shows view-inconsistent flickering in animation, so FID alone misses multi-view consistency.
Limitations
- Relies on parametric SMPL deformation, which does not handle hair and loose clothing dynamics.
- Geometry and normals are hard to extract from irregular sparse Gaussians.
- Results are not yet photorealistic.
Experimental setup
Datasets: DeepFashion SHHQ
Benchmarks: DeepFashion SHHQ
Metrics: FID PCK inference time ms/img
Baselines: EG3D StyleSDF ENARF GNARF StylePeople EVA3D GetAvatar AG3D
Builds on: 3D Gaussian Splatting StyleGAN2 EG3D GNARF Shell Maps
Closest papers by meaning
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.88)
- HumanGaussian: Text-Driven 3D Human Generation with Gaussian Splatting (CVPR 2024 · similarity 0.87)
- GaussianIP: Identity-Preserving Realistic 3D Human Generation via Human-Centric Diffusion Prior (CVPR 2025 · similarity 0.87)
- CGS-GAN: 3D Consistent Gaussian Splatting GANs for High Resolution Human Head Synthesis (NeurIPS 2025 · similarity 0.87)
- GauHuman: Articulated Gaussian Splatting from Monocular Human Videos (CVPR 2024 · similarity 0.86)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (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.