Research map / 3D vision / Face and shape
HumanNOVA: Photorealistic, Universal and Rapid 3D Human Avatar Modeling from a Single Image
Hezhen Hu, Wangbo Zhao, Lanqing Guo, Hanwen Jiang, Jonathan C. Liu, Zhiwen Fan, Kai Wang, Zhangyang Wang, Georgios Pavlakos
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 |
|---|---|---|---|
| add generated data to Real3D fine-tuning | Real3D · CustomHuman (frontal input) | PSNR: 17.13 -> 20.97 | Improved |
| add generated data to Real3D fine-tuning | Real3D · CustomHuman (frontal input) | LPIPS: 95.12 -> 58.54 | Improved |
| add SMPL-X vertex initialization instead of COLMAP | real-world data generation (3DGS fitting) · multi-camera human capture | PSNR: 16.49 -> 36.38 | Improved |
| add SMPL-X vertex initialization instead of COLMAP | real-world data generation (3DGS fitting) · multi-camera human capture | optimization time: 40 minutes to 4 minutes | Improved |
| add the SMPL mesh prior (ablation) | HumanNOVA · CustomHuman (frontal input) | LPIPS: 2.3% relative improvement | Improved |
| use triplane size larger than 32 (ablation) | HumanNOVA · CustomHuman (frontal input) | rendering quality | Improved |
| increase portion of generated training data (ablation) | HumanNOVA · CustomHuman (frontal input) | rendering quality | Improved |
| use Sapiens instead of DINOv2 as visual encoder (ablation) | HumanNOVA · CustomHuman (frontal input) | LPIPS: 42.42 -> 46.52 | No clear effect |
| use 4 fusion layers instead of 2 (ablation) | HumanNOVA · CustomHuman (frontal input) | LPIPS: 50.65 -> 42.42 | Improved |
| use 4 supervision views instead of 2 (ablation) | HumanNOVA · CustomHuman (frontal input) | LPIPS: 45.18 -> 42.42 | Improved |
| train with CustomHuman held out (leave-one-out) (ablation) | HumanNOVA · CustomHuman (frontal input) | LPIPS: 42.42 -> 44.22 | No clear effect |
| use the human-specific HumanNOVA instead of SiTH | single-image avatar reconstruction · CustomHuman (frontal input) | LPIPS: 41.8% relative LPIPS gain | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- SF3D — SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement (CVPR 2025) · better on 6 of 6 shared results
- SiTH — SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion (CVPR 2024) · better on 6 of 6 shared results
- Real3D — Real3D: The Curious Case of Neural Scene Degeneration (AAAI 2024) · better on 2 of 2 shared results
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
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.90)
- IDOL: Instant Photorealistic 3D Human Creation from a Single Image (CVPR 2025 · similarity 0.90)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.89)
- Codec Avatar Studio: Paired Human Captures for Complete, Driveable, and Generalizable Avatars (NeurIPS 2024 · similarity 0.89)
- HumanVid: Demystifying Training Data for Camera-controllable Human Image Animation (NeurIPS 2024 · similarity 0.88)
- DreamAvatar: Text-and-Shape Guided 3D Human Avatar Generation via Diffusion Models (CVPR 2024 · similarity 0.88)
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.
