Source-linked AI summary
HairCS: Reconstructing Strand-Based Hair from Hair Cards
Zixuan Lu, Tongtong Wang, Yuefan Shen, Zhongtian Zheng, Chenfanfu Jiang, Yin Yang, Kui Wu
TL;DR
Manual hair creation is labor-intensive because hairstyles contain rich geometric detail and complex spatial organization. HairCS converts textured hair-card meshes into dense strand-based models that preserve hairstyle structure while producing uniformly rooted, plausibly filled hair volumes compatible with rendering, simulation, and grooming.
Problem
Manual hair creation is labor-intensive because hair has rich geometric detail, complex spatial organization, and distinctive structures across individuals and hairstyles.
Method
HairCS extracts guides from textured hair cards, regularizes their spatial distribution with wrapper-based relaxation, and synthesizes dense strands that preserve the card-defined hairstyle.
Results
The pipeline produces strand-based hair that preserves the original hairstyle while ensuring scalp-rooted strands, uniformly distributed roots, and plausible volumetric filling.
Takeaways & Limitations
The resulting assets support strand-based rendering, simulation, and grooming, and enable conversion of existing card-based content into high-fidelity strand-based hairstyles.
Takeaways & Limitations
The method does not model strand density or cross-sectional variation within wisps and does not guarantee correct topology for braids.
Abstract
from arXiv · showhide
We present an automated pipeline that converts hair-card models into high-quality strand-based hairstyles. Given a collection of textured triangular or quad strips as input, our method produces a strand-based representation that preserves the original hairstyle while enriching it with fine-scale geometric detail and adhering to standard production requirements: strands originate from the scalp, roots are uniformly distributed, and the hair volume is plausibly filled. The resulting assets are directly compatible with strand-based rendering, physics-based simulation, and common grooming modifiers (e.g., clumping, curling, noise) for enhanced realism and artistic control. We validate our approach on a large and diverse set of hairstyles, including short and long hair, curly styles, and complex styles such as buns and ponytails.
1 Introduction
HairCS addresses the costly manual authoring of strand-based hair by converting existing hair-card assets into high-fidelity strand representations. The pipeline preserves hairstyle structure while regularizing roots, filling volume, and supporting production grooming and simulation.
- The method targets conversion of low-cost card assets into strand-based models suitable for AAA game production and large-scale hair datasets.
- HairCS presents an automated pipeline that converts textured hair-card strips into high-quality strand-based hair while preserving the original hairstyle and adding fine-scale geometric detail.
- The resulting strands originate from the scalp, have uniformly distributed roots, plausibly fill the hair volume, and support rendering, simulation, clumping, curling, and fuzz modifiers.
- The pipeline addresses nonuniform card roots, volumetric gaps or crowding, and preservation of partings and layering through guide binding, wrapper relaxation, and density-targeted strand relaxation.
- It extracts guides, redistributes roots over the scalp through integer programming, inflates guides into coverage-constrained wrappers, and instantiates wrappers as strand clusters.
2 Related Work
Related work develops simplified hair representations, manual and procedural modeling tools, image-based generation methods, and increasingly large hairstyle datasets. HairCS builds on this landscape by addressing conversion from widely used hair cards to strand-based models.
- Hair Representation: Hair cards remain widely used in gaming because they provide a simple and efficient reduced hair representation that can be expanded during rendering.
- Hair Modeling: Traditional hair authoring remains labor-intensive, motivating volumetric mesh and geometric methods that provide higher-level editing or target specialized hairstyles.
- Hair Generation: Hair generation research includes sketch-based interfaces and automatic strand reconstruction from images using heuristics, hairstyle databases, and learning-based approaches.
- Hair Dataset: USC-HairSalon established an influential public 3D hairstyle dataset, while newer resources expand scale, quality, and stylistic diversity for data-driven hair research.
3 Overview
HairCS formulates hair-card conversion as producing dense strands that preserve card structure while satisfying uniform scalp coverage and plausible volumetric filling. Its pipeline extracts and augments guides, relaxes wrapper volumes, and refines the final strands.
- Problem Statement: The task is to generate dense strand polylines from textured hair cards, with roots on the scalp, uniform root distribution, preserved hairstyle structure, and plausible volume filling.
- Problem Statement: The central difficulty is that cards emphasize external appearance, have nonuniform scalp placement, and represent flat cluster proxies rather than volumetric curves.
- Our Pipeline: The pipeline extracts guides, binds their roots to uniformly sampled scalp positions, and adds guides traced from farthest-point-sampled unbound scalp locations.
- Our Pipeline: A subsequent strand-level relaxation removes gaps and discontinuities between wrapper-generated clusters while remaining inside the hair volume and preserving the initial hairstyle shape.
4 Method
The method extracts oriented guides from hair cards, binds them to scalp roots while respecting hair partings, fills uncovered regions with additional guides, and regularizes the resulting guide distribution within a reconstructed hair volume.
- 4.1 Guide Extraction: The method extracts each card’s dominant flow orientation, root-to-tip direction, and evenly spaced guides from its texture, then lifts them into 3D using the card’s UV mapping.It assumes each card is a UV strip with a monotone flow axis, but does not require fixed UV orientation, root side, shared-texture exclusivity, or flat geometry.
- 4.2 Guide Binding: Guide roots are assigned to scalp candidates by minimizing distance and angular inconsistency, with each guide receiving one root and each candidate serving at most one guide.The assignment is solved as a rectangular linear sum assignment problem using the Hungarian algorithm.
- 4.2 Guide Binding: Parting-aware reassignment detects opposing projected guide orientations across a scalp Voronoi tessellation, fits partition curves through their boundary candidates, and moves guides crossing to the correct side.This addresses crossings that arise when binding uses only geometric proximity.
- 4.3 Root Tracing: Additional guides are selected by farthest-point sampling from unassigned roots to cover under-sampled scalp regions, excluding roots near partition curves before tracing.Each selected root is traced along the scalp normal to the nearest card and then along that card’s flow direction when a card is hit.
- 4.3 Root Tracing: The full guide set is smoothed with a root-normality and curvature-regularized optimization, then represented as tube-shaped wrappers whose admissible volume combines rasterized card volume with gaps to the bust.A target density diffuses scalp-root density through this volume, while one-sided signed-distance penalties constrain the relaxation.
5 Results
The method produces faithful strand-based hairstyles across diverse assets while preserving structure, filling volume, and meeting production constraints. Ablations and comparisons show that distributed guides, relaxation, and wrapper-based synthesis are central to quality.
- 5.1 Evaluations: The pipeline runs in about 1–2 minutes for approximately 100K dense strands on hundreds of cards, using less than 500MB of memory.Full-hair synthesis is the most time-consuming stage because it directly manipulates dense strands.
- 5.1 Evaluations: The method reproduces short, long, wavy, ponytail, and bun hairstyles, including approximately three hundred generated strand-based examples.Results are produced with one parameter set across assets, except for increased sampling on strongly curved in-the-wild cards.
- 5.1 Evaluations: With 200 extra guides, strands become smooth and uniformly distributed; without them, empty regions remain and both CD and d_c→s degrade.Beyond a small guide budget, the two geometric metrics stabilize while root coverage continues improving.
- 5.2 Comparison with Alternative Hair Synthesis Methods: Our wrapper-based interpolation preserves within-wisp features, while density relaxation resolves clumpiness and discontinuities between wrappers without manual post-editing.Prism interpolation causes penetration and unwanted strands, whereas clump interpolation creates boundary artifacts.
- 5.4 Texture-Guided Initialization: Texture-guided tracing follows high-fidelity flow information, but most evaluated assets provide only RGB textures, leaving matching rendered results for future work.The texture-guided route is substantially slower on the demonstrated asset.
6 Applications
The strand output supports applications beyond reconstruction, including in-the-wild conversion, texture-guided initialization, hairstyle blending, grooming, and simulation. These applications use scalp connectivity, guide structure, and volumetric wrappers to preserve coherent hair behavior.
- 6 Applications: The pipeline remains robust on hundreds to thousands of complex, strongly curved cards from additional in-the-wild asset packages.These assets use 128 samples per strand because of their curvature.
- 6.2 Texture-Guided Initialization: Texture-guided initialization combines ID and flow maps to trace streamlines along hair direction before lifting them into 3D.The method uses this route when high-fidelity texture maps are available.
- 6.1 Hairstyle Blending: Scalp-rooted, guide-traceable strands enable semantic template blending by replacing selected scalp regions between converted hairstyles.The process partitions the scalp into semantic regions, filters orientation-incompatible imported guides, and preserves downstream processing.
- 6.3 Grooming: Wrapper-based output supports bend, scale, fuzz, curly, and helical grooming while preserving each wisp’s volumetric distribution.The resulting strands are compatible with conventional grooming systems such as Maya XGen and Houdini.
- 6.4 Simulation: Because strands are scalp-connected, the output can be sent directly to a strand simulator and rendered with real-time wind excitation in Unreal Engine.The demonstrated simulation operates without additional processing.
7 Conclusion and Limitations
The paper presents a five-step pipeline that converts hair cards into high-fidelity strand models, supports grooming and simulation, and enables a 50K diverse hairstyle dataset. Its limitations include unresolved within-wisp strand distribution, hybrid card-and-mesh assets, and braid topology.
- 7 Conclusion and Limitations: The five-step pipeline converts hair cards into high-fidelity strand models and bridges low-cost card assets with simulation-, dataset-, and data-driven applications.The method addresses volumetric occupancy and layered hairstyle structure through wrapper-based relaxation, dense synthesis, and subsequent refinement.
- 7 Conclusion and Limitations: 50K diverse hairstyles are enabled by combining the conversion pipeline with hairstyle blending and grooming modifiers.The resulting strand-based hair can be enriched with curls, waves, and helical structures.
- 7 Conclusion and Limitations: The method does not model strand distribution within a wisp, including density and cross-sectional variation along the centerline.Reconstructing strands directly from card textures remains an open challenge.
- 7 Conclusion and Limitations: Hybrid card-and-mesh assets cannot achieve the promised quality because correct centerline extraction depends on the cards’ 2D parameterization.The authors identify reconstruction of strands for such hybrid assets as a next step.
- 7 Conclusion and Limitations: The method does not guarantee correct topology for braids because guides are extracted directly from the cards.
A Wrapper Relaxation Algorithm
Wrapper relaxation expands guide-associated wrappers while constraining their geometry and uses an XPBD rod solver to relax guide strands under wrapper collisions.
- A Wrapper Relaxation Algorithm: The wrapper set forms an approximate partition of the generated hair volume.
- A Wrapper Relaxation Algorithm: Guide strands are relaxed under wrapper-induced collisions, with blocked cross-sectional displacement driving guide dynamics.The guides are modeled as elastic rods and solved with XPBD after iterative ACCD-clamped wrapper expansion.
- A Wrapper Relaxation Algorithm: The optimization energy combines momentum, segment-length constraints, and rest-angle constraints.Stiffness values and sweep counts are specified in Table 3, while strands are processed independently and in parallel.
B Wrapper-based Hair Deformation
Wrapper-based deformation maps scalp-root coordinates along relaxed guide frames into wrapper cross-sections, producing dense strands that remain inside their assigned volumes.
- B Wrapper-based Hair Deformation: Dense synthesis converts each guide, wrapper, and scalp region into a scalp-connected wisp whose strands fill the wrapper uniformly and remain inside it.
- B Wrapper-based Hair Deformation: The method avoids naive root translation by transporting canonical disk coordinates and fitting them radially to each wrapper cross-section.This handles non-circular and spatially varying wrapper cross-sections.
- B Wrapper-based Hair Deformation: Axis-angular coordinates represent points inside a wrapper using arc length, angular position, and normalized radius.The normalized radial coordinate reaches the wrapper boundary at η = 1.
- B Wrapper-based Hair Deformation: Each scalp root receives canonical coordinates that are parallel-transported along its guide before continuous mapping generates strand samples inside the wrapper.
- B Wrapper-based Hair Deformation: The deformation produces scalp-connected strands that conform to local wrapper geometry, with Fig. 30 showing plausible results across challenging guide configurations.
C Helix Operator with Rotational Transport
The helix operator creates curly or wavy guides before strand synthesis, transports strand offsets along the deformed guides, and then conforms the bundle to wrapper geometry.
- C Helix Operator with Rotational Transport: Curly synthesis modifies guide initialization directly rather than applying a post-hoc operator to straight guides.Orientation-field construction, density-field construction, and relaxation remain unchanged afterward.
- C Helix Operator with Rotational Transport: The operator exposes turn count, ramp exponent, and helix radius to control rotational accumulation and lateral amplitude along each guide.Normalized arc length runs from scalp root to tip, and the angular displacement is defined from these parameters.
- C Helix Operator with Rotational Transport: Rotational transport uses incremental Rodrigues rotations to move each strand’s cross-sectional offset along the deformed guide.The procedure preserves a constant signed angle relative to the initial frame while the frame follows the tangent.
- C Helix Operator with Rotational Transport: The parameter set is fixed across every reported result, with head dimensions normalized to unit scale.
- C Helix Operator with Rotational Transport: The transported bundle is warped through per-cross-section deformation so helix-mode strands remain inside and match the deformed wrapper.
D Interpolation Settings
The interpolation settings define two Houdini Hair Generate baselines that share common parameters but differ in how guides generate strands.
- The two baselines use the same guide set and are hand-tuned with the settings listed in Table 4.Generated strands are not edited after interpolation.
- The prism-based baseline blends up to three guides linearly.
- The clump-based baseline follows a single guide by extrusion.
E Hyperparameters
The section specifies a shared parameter set and organizes the dataset by hairstyle production mode, length class, and helix application.
- All results use the parameter set reported in Table 3.
- Table 5 separates hairstyles into reconstructed or blended production modes, each provided in straight and helix variants.
- The dataset composition is further divided by five length/type labels: short, bob, shoulder, long, and gather.
F 50K Strand-based Hair Dataset
HairCS provides a large strand-based hairstyle dataset with shared geometry, standardized strand arrays, multiple production variants, and supporting visualization tools.
- Each hairstyle uses the same bust and scalp coordinate frame, with 60,000 strands resampled to 64 root-to-tip points.Roots are uniformly scattered over the scalp region, and coordinates are stored in float16 arrays.
- The dataset contains 53,336 hairstyles, including reconstructed outputs, grooming-operator variants, and blended hairstyles across five length/type classes.The dataset is distributed through Hugging Face, and Table 5 documents its composition.
- Only 4,464 hairstyles are direct pipeline outputs, while 48,872 come from hairstyle blending toward short, bob, shoulder, and long targets.
- Each hairstyle includes a rendered preview image, and the repository provides a lightweight strand viewer.
- A strand-to-surface tool converts hairstyles into watertight meshes using voxel rasterization, signed distance fields, marching cubes, and isotropic remeshing.