Research map / 3D vision / Face and shape
URHand: Universal Relightable Hands
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 URHand, a universal relightable hand model trained on multi-identity light-stage data with a spatially varying linear lighting model and hybrid neural-physical branches, enabling generalization to novel identities, poses and illuminations and few-shot phone-scan personalization.
Problem
Photorealistic relightable hand models need identity-specific multi-view, multi-illumination captures and do not generalize to novel identities or natural illuminations; neural relighting needs costly teacher-student distillation to handle continuous environment lighting.
Approach
A spatially varying linear neural renderer (non-linear activations and biases removed, gated by non-linear pose/identity branch) takes physically based shading features as input, preserving light-transport linearity so that training on grouped point lights generalizes to environment maps without distillation; combined with a physical branch for geometry refinement and a lighting-aware adversarial
Key results
- On per-subject training, the full model reaches 27.77 PSNR on Subject 1 versus 25.97 for RelightableHands.
- On unseen identities, the full model achieves 26.94 PSNR versus 16.65 for Handy + Phong on Subject 1.
- Runtime of 38 FPS with grouped lights and 31 FPS with environment maps on an NVIDIA A100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use URHand instead of RelightableHands | Hand relighting · Subject 1, unseen segments | PSNR: 25.97 -> 27.77 | Helped |
| Use URHand instead of Handy + Phong | Hand relighting · Subject 1, unseen identity | PSNR: 16.65 -> 26.94 | Helped |
| Use a spatially varying linear lighting model rather than non-linear activations (ablation) | URHand · Fully lit | PSNR: 17.50 -> 20.61 | Helped |
| Use a spatially varying linear lighting model rather than the MLP-based linear model (ablation) | URHand · Unseen subject | PSNR: 21.56 -> 24.70 | Helped |
| Add the specular lighting feature (ablation) | URHand · Unseen subject | PSNR: 23.20 -> 24.70 | Helped |
| Add visibility to the lighting features (ablation) | URHand · Unseen subject | PSNR: 23.75 -> 24.70 | Helped |
| Add the geometry refiner (ablation) | URHand · Unseen subject | PSNR: 22.80 -> 24.70 | Helped |
| Add the adversarial loss (ablation) | URHand · Unseen subject | PSNR: 23.70 -> 24.70 | Helped |
| Add lighting awareness to the adversarial loss (ablation) | URHand · Trained subject | PSNR: 25.38 -> 26.01 | Helped |
| Add L1 regularization to the linear model (ablation) | URHand · Unseen subject | PSNR: 23.90 -> 24.70 | Helped |
Surprising findings
- A holistic MLP-based linear lighting model (Yang et al.) performs poorly on hands, with 11.56 PSNR on fully lit frames, due to shadow variations from articulation.
Limitations
- Personalization preprocessing (geometry fitting and texture unwrapping) takes several hours.
- Quality slightly degrades for unseen identities compared with per-identity optimization.
Experimental setup
Datasets: light-stage multiview hand captures (93 identities)
Benchmarks: held-out light-stage hand sequences (single-identity and novel identity)
Metrics: PSNR SSIM LPIPS
Baselines: RelightableHands Handy MLP-based linear model (Yang et al.) Phong relighting
Builds on: RelightableHands Disney BRDF UHM Yang et al. linear face relighting
Closest papers by meaning
- FoundHand: Large-Scale Domain-Specific Learning for Controllable Hand Image Generation (CVPR 2025 · similarity 0.86)
- 3D Visibility-Aware Generalizable Neural Radiance Fields for Interacting Hands (AAAI 2024 · similarity 0.86)
- OHTA: One-shot Hand Avatar via Data-driven Implicit Priors (CVPR 2024 · similarity 0.86)
- HanDiffuser: Text-to-Image Generation With Realistic Hand Appearances (CVPR 2024 · similarity 0.85)
- HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing (NeurIPS 2026 · similarity 0.85)
- SRHand: Super-Resolving Hand Images and 3D Shapes via View/Pose-aware Neural Image Representations and Explicit Meshes (NeurIPS 2025 · similarity 0.84)
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.