Research map / 3D vision / Face and shape
GroomLight: Hybrid Inverse Rendering for Relightable Human Hair Appearance Modeling
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.
GroomLight combines an extended analytical hair BSDF (scalp-aligned albedo and rotation correctives) with a light-aware residual of Gaussian dual-level spherical harmonics in a two-stage hybrid inverse rendering pipeline for relightable hair from multi-view OLAT images.
Problem
Relightable photorealistic human hair appearance modeling from multi-view OLAT captures: analytical BSDF models miss appearance details, while neural rendering overfits and relights poorly.
Approach
Pair an extended Chiang-style hair BSDF (spatially varying scalp-aligned absorption, rotation correctives for tangent errors) that captures primary light transport with a strand-aligned 3D Gaussian residual whose color is a light- and view-conditioned dual-level SH, optimized in two stages with differentiable path tracing and then Gaussian rasterization.
Key results
- GroomLight reaches 29.71 PSNR, 0.7727 SSIM, 0.2461 LPIPS on novel-view reconstruction, versus 27.22 PSNR for GaussianHair and 26.50 for HairInverse.
- The residual model raises PSNR from 28.67 (Stage1 Full) to 29.71, showing the residual's contribution.
- Stage-1 extended BSDF (spatially varying albedo plus rotation correctives) improves PSNR from 27.75 to 28.67 over the original BSDF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GroomLight's full hybrid appearance model instead of GaussianHair | GroomLight · Novel-view reconstruction | PSNR: 27.22 -> 29.71 | Helped |
| Use GroomLight's full hybrid appearance model instead of GaussianHair | GroomLight · Novel-view reconstruction | SSIM: 0.7340 -> 0.7727 | Helped |
| Use GroomLight's full hybrid appearance model instead of GaussianHair | GroomLight · Novel-view reconstruction | LPIPS: 0.2920 -> 0.2461 | Helped |
| Add spatially varying albedo to the original hair BSDF (ablation) | GroomLight · Novel-view reconstruction | PSNR: 27.75 -> 28.58 | Helped |
| Add rotation correctives to the spatially varying albedo model (ablation) | GroomLight · Novel-view reconstruction | PSNR: 28.58 -> 28.67 | Helped |
| Add the light-aware residual model to the full Stage 1 model (ablation) | GroomLight · Novel-view reconstruction | PSNR: 28.67 -> 29.71 | Helped |
Limitations
- Requires light stage capture for lighting and camera calibration, so it does not apply to in-the-wild data.
- Offline optimization renders at about 1 FPS, not real-time.
- Needs accurate geometry and segmentation masks to avoid artifacts from clothing, face, or background.
Experimental setup
Datasets: multi-view OLAT light stage captures of 10 subjects
Benchmarks: in-house 10-subject light-stage hair dataset (novel views and 160 novel lightings)
Metrics: PSNR SSIM LPIPS
Baselines: HairInverse GaussianHair
Builds on: Chiang et al. hair BSDF Mitsuba 3 3D Gaussian Splatting GroomCap path replay backpropagation
Closest papers by meaning
- SynthLight: Portrait Relighting with Diffusion Model by Learning to Re-render Synthetic Faces (CVPR 2025 · similarity 0.87)
- MonoHair: High-Fidelity Hair Modeling from a Monocular Video (CVPR 2024 · similarity 0.87)
- Relightable Gaussian Codec Avatars (CVPR 2024 · similarity 0.87)
- Perm: A Parametric Representation for Multi-Style 3D Hair Modeling (ICLR 2025 · similarity 0.87)
- Dr.Hair: Reconstructing Scalp-Connected Hair Strands without Pre-Training via Differentiable Rendering of Line Segments (CVPR 2024 · similarity 0.86)
- BecomingLit: Relightable Gaussian Avatars with Hybrid Neural Shading (NeurIPS 2025 · 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.