Research map / 3D vision / Face and shape
Quaffure: Real-Time Quasi-Static Neural Hair Simulation
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 Quaffure, a self-supervised neural network that predicts physically plausible quasi-static hair drapes conditioned on hairstyle latent, body shape and pose, running in milliseconds without simulated training data.
Problem
Real-time simulation of quasi-static hair drape across body poses, shapes and hairstyles is computationally expensive for physics-based methods, and supervised learned simulators need costly simulated training data.
Approach
Decomposes hair deformation into a rigid pose-based groom transformation plus a learned deformation decoder trained only with a physics-based self-supervised loss (modified Cosserat stretch-shear energy, gravity, body and self-collision, pose regularization), with hairstyles encoded by a pretrained autoencoder latent on a scalp UV texture.
Key results
- Single groom inference takes 2.86 ms and 1000 grooms take 0.3 s on an RTX A6000.
- Compared with GroomGen, Ours reports 0.26% body intersection versus 0.39% and length error 175.42 versus 1319.74.
- Achieves metrics comparable to Adam/L-BFGS energy minimization (e.g., 0.26% vs 0.22% intersection) at orders of magnitude lower time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the modified Cosserat energy instead of mass-spring energy (ablation) | Quaffure · Hair shape and curliness preservation | hair shape preservation | Helped |
| Increase Cosserat stiffness (ablation) | Quaffure · Hair strand rest-shape preservation | rest-shape preservation | Helped |
| Use Quaffure instead of GroomGen | Quasi-static hair draping · Quasi-static hair draping | body intersection (%): 0.39 -> 0.26 | Helped |
| Use Quaffure instead of GroomGen | Quasi-static hair draping · Quasi-static hair draping | length preservation: 1319.74 -> 175.42 | Helped |
| Use Quaffure instead of GroomGen | Quasi-static hair draping · Quasi-static hair draping | orientation preservation: 1281.96 -> 286.13 | Helped |
| Use Quaffure instead of GroomGen | Quasi-static hair draping · Quasi-static hair draping | time (seconds): 0.00249 -> 0.00286 | Hurt |
Surprising findings
- Full Cosserat rod model with orientation offsets was infeasible due to slow training and convergence, motivating a position-only modified Cosserat energy.
Limitations
- Pose-based transformation can initialize strands partially inside the body.
- Uses a single set of physical material parameters; not conditioned on material.
- Implausible deformations for non-physical poses beyond training data (e.g., neck bent unnaturally).
Experimental setup
Datasets: CT-groom motion capture pose dataset
Benchmarks: in-house groom and pose evaluation
Metrics: inference time % body intersection length preservation orientation preservation gravity potential
Baselines: Adam L-BFGS XPBD GroomGen
Builds on: Cosserat rod model XPBD GroomGen SPH density estimation
Closest papers by meaning
- DGH: Dynamic Gaussian Hair (NeurIPS 2025 · similarity 0.88)
- HairFastGAN: Realistic and Robust Hair Transfer with a Fast Encoder-Based Approach (NeurIPS 2024 · similarity 0.86)
- Perm: A Parametric Representation for Multi-Style 3D Hair Modeling (ICLR 2025 · similarity 0.86)
- SimAvatar: Simulation-Ready Avatars with Layered Hair and Clothing (CVPR 2025 · similarity 0.85)
- Doubly Hierarchical Geometric Representations for Strand-based Human Hairstyle Generation (NeurIPS 2024 · similarity 0.85)
- LUCAS: Layered Universal Codec Avatars (CVPR 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.