Research map / 3D vision / 3D generation
Joint2Human: High-Quality 3D Human Generation via Compact Spherical Embedding of 3D Joints
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.
Joint2Human generates detailed, high-resolution 3D human geometry by running a latent diffusion model directly on Fourier occupancy fields, with compact spherical joint embeddings for pose control, a high-frequency enhancer, and multi-view recarving.
Problem
Generating 3D human geometry that simultaneously has global structural consistency, fine local details, high resolution, and low computational cost; 2D-lifting methods lack view consistency and native 3D methods (tri-planes, SDFs) are costly and lose detail.
Approach
Operate latent diffusion directly on image-aligned Fourier occupancy fields (FOF) so 2D diffusion yields 3D shapes natively, guide pose with a compact spherical embedding of SMPL's 24 3D joints (rendered as spheres in FOF channels), recover high-frequency terms with a skip-connected enhancer decoder, and fuse two orthogonal views by blending occupancy fields.
Key results
- Ours (fine) achieves FID_normal 23.89 and FID_shade 41.20 on CustomHumans, versus Chupa fine at 29.90 and 45.49.
- In a user study with 123 responses, 49.10% preferred Ours for global structure versus 32.20% for Chupa.
- The method generates a human in under one minute, per the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Joint2Human fine generation instead of Chupa fine | 3D human generation · CustomHumans test | FIDnormal: 29.90 -> 23.89 | Helped |
| use Joint2Human fine generation instead of Chupa fine | 3D human generation · CustomHumans test | FIDshade: 45.49 -> 41.20 | Helped |
| use CSEJ instead of SMPL for pose guidance (ablation) | Joint2Human · CustomHumans test | FIDnormal: 49.16 -> 45.73 | Helped |
| add CSEJ pose guidance to IUV-map pose guidance (ablation) | Joint2Human · CustomHumans test | FIDnormal: 39.51 -> 38.40 | Helped |
| use CSEJ alone instead of IUV-map pose guidance (ablation) | Joint2Human · CustomHumans test | FIDnormal: 39.51 -> 45.73 | Hurt |
| add multi-view recarving (ablation) | Joint2Human · CustomHumans test | FIDnormal: 37.39 -> 31.90 | Helped |
| add the high-frequency enhancer (ablation) | Joint2Human · CustomHumans test | FIDnormal: 37.39 -> 29.51 | Helped |
| add the high-frequency enhancer when recarving is enabled (ablation) | Joint2Human · CustomHumans test | FIDnormal: 31.90 -> 25.67 | Helped |
| add multi-view recarving when the high-frequency enhancer is enabled (ablation) | Joint2Human · CustomHumans test | FIDnormal: 29.51 -> 25.67 | Helped |
Surprising findings
- Conditioning on SMPL parameters led to overfitting with FOF representations, motivating the joint-based embedding.
Limitations
- Extreme poses such as stooping or standing on the head remain challenging.
- Multi-view recarving is performed only once (theta = pi/2), trading off quality for efficiency.
Experimental setup
Datasets: THuman 2.0 THuman 3.0 2K2K commercial human meshes (about 1500) CustomHumans (test)
Benchmarks: CustomHumans
Metrics: FID normal FID shade user preference
Baselines: Chupa AG3D
Builds on: latent diffusion model Chupa Fourier occupancy field DensePose Text2Human
Closest papers by meaning
- SyncHuman: Synchronizing 2D and 3D Generative Models for Single-view Human Reconstruction (NeurIPS 2025 · similarity 0.89)
- HumanNorm: Learning Normal Diffusion Model for High-quality and Realistic 3D Human Generation (CVPR 2024 · similarity 0.87)
- PSHuman: Photorealistic Single-image 3D Human Reconstruction using Cross-Scale Multiview Diffusion and Explicit Remeshing (CVPR 2025 · similarity 0.87)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.87)
- HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion (ICLR 2024 · similarity 0.87)
- Human-3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models (NeurIPS 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.