Research map / 3D vision / Gaussian splatting
DGH: Dynamic Gaussian Hair
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 Dynamic Gaussian Hair (DGH), a data-driven coarse-to-fine network that predicts head-motion-driven hair deformation and a strand-guided dynamic 3D Gaussian appearance model, trained on a new synthetic XPBD-simulated hair dataset.
Problem
Photorealistic dynamic hair modeling and rendering under head motion is hard: physics-based simulation needs per-hairstyle parameter tuning and is costly, and existing Gaussian hair methods handle only static or rigidly transformed hair.
Approach
A volumetric SDF-based feed-forward network predicts time-independent pose-driven displacements (with a hair-to-body SDF collision penalty) and then temporal flow refinement via latent cross-attention; appearance uses cylindrical Gaussians per strand segment with an MLP conditioned on tangents and curvature-based blending of SH and opacity.
Key results
- Appearance: average PSNR 27.01, SSIM 0.871, LPIPS 0.127 versus Gaussian Haircut 23.07/0.847/0.131 and 3DGS 20.91/0.822/0.164 across 5 subjects.
- Deformation: average L2 error 0.0832 and Chamfer 0.0266 versus 0.1639 and 0.0424 for rigid-transformed hair.
- Inference on a single A100 reports 2.0 FPS for about 150K strands versus about 0.33 FPS for XPBD simulation; rendering 2.22 FPS versus 0.17 FPS for Blender shader.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DGH instead of Gaussian Haircut | Hair appearance rendering · Hair appearance, Average | PSNR: 23.073 -> 27.009 | Helped |
| Use DGH instead of 3D GS | Hair appearance rendering · Hair appearance, Average | SSIM: 0.822 -> 0.871 | Helped |
| Use DGH instead of rigged hair | Hair deformation · Hair deformation, Average | Error (L2): 0.1639 -> 0.0832 | Helped |
| Add the hair-to-body SDF constraint (ablation) | DGH hair deformation · Hair deformation test sets | Error (L2): 0.1269 -> 0.0832 | Helped |
| Add the fine-stage motion model (ablation) | DGH hair deformation · Hair deformation test sets | Error (L2): 0.0964 -> 0.0832 | Helped |
| Add cross-attention to the fine stage (ablation) | DGH hair deformation · Hair deformation test sets | Error (L2): 0.0909 -> 0.0832 | Helped |
| Add tangent features and curvature-based blending (ablation) | DGH hair appearance · Hair appearance | PSNR: 20.89 -> 28.12 | Helped |
| Add curvature-based blending (ablation) | DGH hair appearance · Hair appearance | PSNR: 25.08 -> 28.12 | Helped |
| Add tangent features and curvature-based blending (ablation) | DGH hair appearance · Hair appearance | LPIPS: 0.19 -> 0.19 | No clear effect |
| Add curvature-based blending (ablation) | DGH hair appearance · Hair appearance | LPIPS: 0.18 -> 0.19 | Hurt |
| Add LPIPS and SSIM losses (ablation) | DGH appearance optimization · Hair appearance optimization | LPIPS: 0.2100 -> 0.1881 | Helped |
| Add LPIPS and SSIM losses (ablation) | DGH appearance optimization · Hair appearance optimization | SSIM: 0.8874 -> 0.9004 | Helped |
Surprising findings
- The paper's own reported results are internally inconsistent in places: the appearance ablation reports PSNR 28.12 for the full model while the main table averages 27.01, and the 'w/o blend' variant has better LPIPS than the full model.
Limitations
- Trained and evaluated only on synthetic data; no real hair tracking data.
- Only handles hair-upper-body collisions; no environmental constraints or relighting.
- Dense hair (100K-150K strands) is not real-time for rendering; full-body effects such as jumping are not modeled.
Experimental setup
Datasets: Synthetic dynamic hair dataset (XPBD simulation with motion-captured head motion, Blender multi-view rendering) Mixamo MoCap data
Benchmarks: synthetic dynamic hair dataset (5 subjects)
Metrics: PSNR SSIM LPIPS L2 error Chamfer distance flow error
Baselines: 3D Gaussian Splatting Gaussian Haircut Rigged hair XPBD
Builds on: 3D Gaussian Splatting Gaussian Haircut XPBD NeRF
Closest papers by meaning
- Perm: A Parametric Representation for Multi-Style 3D Hair Modeling (ICLR 2025 · similarity 0.89)
- MeGA: Hybrid Mesh-Gaussian Head Avatar for High-Fidelity Rendering and Head Editing (CVPR 2025 · similarity 0.89)
- LUCAS: Layered Universal Codec Avatars (CVPR 2025 · similarity 0.88)
- Doubly Hierarchical Geometric Representations for Strand-based Human Hairstyle Generation (NeurIPS 2024 · similarity 0.88)
- Quaffure: Real-Time Quasi-Static Neural Hair Simulation (CVPR 2025 · similarity 0.88)
- DiffLocks: Generating 3D Hair from a Single Image using Diffusion Models (CVPR 2025 · 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.