Research map / Image generation / Editing and style transfer
OHTA: One-shot Hand Avatar via Data-driven Implicit Priors
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 OHTA, a two-stage method that learns a Hand Prior Network (geometry, albedo, shadow fields on a MANO-HD mesh scaffold) from multi-identity data and then inverts and fits it to create a drivable implicit hand avatar from a single image.
Problem
Creating high-fidelity, animatable hand avatars from only one image, where existing implicit hand avatar methods need many sequential or multi-view images.
Approach
A Hand Prior Network with a mesh-guided multi-resolution field and an identity-shared shadow field (self-occlusion prior) plus identity-specific albedo is pre-trained on multi-identity data, then texture inversion (identity code and color calibration) and partial texture fitting with view regularization adapt it to a single target image.
Key results
- On InterHand2.6M one-shot, OHTA reaches 26.11 PSNR / 12.93 LPIPS / 0.864 SSIM versus Handy's 25.56 / 14.98 / 0.794 and HandAvatar's one-shot 23.79 / 17.78 / 0.820.
- On HanCo, OHTA achieves 22.15 PSNR and 0.886 SSIM, outperforming Handy (21.11) and HandAvatar (19.76).
- Ablation shows texture inversion and view regularization are important; removing inversion drops PSNR from 26.11 to 25.47.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use OHTA instead of HandAvatar | one-shot hand avatar reconstruction · InterHand2.6M | PSNR: 28.23 -> 26.11 | Hurt |
| use OHTA instead of HandAvatar | one-shot hand avatar reconstruction · InterHand2.6M | LPIPS: 10.35 -> 12.93 | Hurt |
| use OHTA instead of HandAvatar | one-shot hand avatar reconstruction · HanCo | PSNR: 19.76 -> 22.15 | Helped |
| use OHTA instead of Handy | one-shot hand avatar reconstruction · HanCo | LPIPS: 11.71 -> 11.55 | Helped |
| add the multi-resolution field (ablation) | HPNet · InterHand2.6M, 7 unseen poses | PSNR: 27.11 -> 27.64 | Helped |
| add the multi-resolution field (ablation) | HPNet · InterHand2.6M, 7 unseen poses | LPIPS: 12.96 -> 12.23 | Helped |
| add the shadow field (ablation) | HPNet · InterHand2.6M, 7 unseen poses | PSNR: 27.09 -> 27.64 | Helped |
| add shape fitting (ablation) | HPNet · InterHand2.6M, 7 unseen poses | PSNR: 27.27 -> 27.64 | Helped |
| add the texture prior (ablation) | OHTA one-shot reconstruction · InterHand2.6M | LPIPS: 14.35 -> 12.93 | Helped |
| add texture fitting (ablation) | OHTA one-shot reconstruction · InterHand2.6M | PSNR: 25.87 -> 26.11 | Helped |
| add texture inversion (ablation) | OHTA one-shot reconstruction · InterHand2.6M | PSNR: 25.47 -> 26.11 | Helped |
| add color calibration (ablation) | OHTA one-shot reconstruction · InterHand2.6M | PSNR: 26.05 -> 26.11 | Helped |
Surprising findings
- One-shot OHTA performs comparably to HandAvatar trained on ~11,757 multi-view images in PSNR (26.11 vs 28.23 is lower, but LPIPS 12.93 vs 10.35 shows a gap), so the gap to full-data methods is modest rather than large.
Limitations
- Suboptimal performance for images with notably uneven lighting.
- Degraded results when pose estimation from the input image is highly inaccurate.
- Relies on an off-the-shelf hand pose estimator and MANO-based shape parameters rather than a refined mesh from one view.
Experimental setup
Datasets: InterHand2.6M HanCo MSCOCO
Benchmarks: InterHand2.6M HanCo
Metrics: PSNR LPIPS SSIM
Baselines: HandAvatar HARP LiveHand Handy SelfRecon HumanNeRF LISA One-2-3-45
Builds on: HandAvatar PairOF NeRF MANO PhoneScan
Closest papers by meaning
- Learning Interaction-aware 3D Gaussian Splatting for One-shot Hand Avatars (NeurIPS 2024 · similarity 0.88)
- HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing (NeurIPS 2026 · similarity 0.87)
- HanDiffuser: Text-to-Image Generation With Realistic Hand Appearances (CVPR 2024 · similarity 0.86)
- FoundHand: Large-Scale Domain-Specific Learning for Controllable Hand Image Generation (CVPR 2025 · similarity 0.86)
- VTON-HandFit: Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding (CVPR 2025 · similarity 0.86)
- URHand: Universal Relightable Hands (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.