Research map / 3D vision / 3D generation
Generative Human Geometry Distribution
Xiangjun Tang, Biao Zhang, Peter Wonka
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SMPL-based training pairs and distribution normalization instead of Gaussian source (ablation) | Human geometry distribution · THuman2 dataset fitting | Chamfer distance: 0.0101 -> 0.0032 | Improved |
| add nearest-point training pair construction (ablation) | Human geometry distribution · THuman2 dataset fitting | Chamfer distance: 0.0706 -> 0.0032 | Improved |
| add distribution normalization (zero-centered displacement field) (ablation) | Human geometry distribution · THuman2 dataset fitting | Chamfer distance: 0.0071 -> 0.0032 | Improved |
| add distribution normalization (ablation) | Human geometry distribution · single geometry fitting | Chamfer distance: 0.0020 -> 0.0032 | Hurt |
| sample SMPL points sparsely before nearest-point search instead of using the dense mesh (ablation) | Training pair construction · THuman2 normal maps | sampling completeness (qualitative) | Improved |
| use auto-decoder instead of VecSet auto-encoder (ablation) | Latent feature map encoding · THuman2 | Surface distance: 0.0018 -> 0.0012 | Improved |
| use a 2D feature map latent instead of independent embeddings in the auto-encoder (ablation) | VecSet auto-encoder · THuman2 | Surface distance: 0.0018 -> 0.0014 | Improved |
| generate raw geometry directly with geometry distributions instead of rendering-enhanced baselines | Human generation · THuman2 pose-conditioned generation | FID: 42.90 -> 16.16 (raw); 17.43 -> 16.16 (vs enhanced gDNA) | Improved |
| synthesize pose-aware feature maps for novel poses | Novel pose generation · user study (25 participants) | Plausibility (1-5): 2.66 -> 4.36 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- gDNA — gDNA: Towards Generative Detailed Neural Avatars (CVPR 2022) · better on 3 of 3 shared results
- EVA3D — EVA3D: Compositional 3D Human Generation from 2D Image Collections (ICLR 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Human Geometry Distribution for 3D Animation Generation (CVPR 2026 · similarity 0.94)
- Human-3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models (NeurIPS 2024 · similarity 0.89)
- SCULPT: Shape-Conditioned Unpaired Learning of Pose-dependent Clothed and Textured Human Meshes (CVPR 2024 · similarity 0.89)
- HumanNorm: Learning Normal Diffusion Model for High-quality and Realistic 3D Human Generation (CVPR 2024 · similarity 0.88)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.87)
- Joint2Human: High-Quality 3D Human Generation via Compact Spherical Embedding of 3D Joints (CVPR 2024 · similarity 0.87)
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.
