Modeling 3D hair strands is challenging due to complex hairstyle geometry, scarce high-quality training data, and incomplete or noisy input observations. Foundation models offer broader hairstyle priors, but their outputs are not directly usable for strand-based rendering and simulation. Existing approaches require costly per-instance optimization to fit hair priors to meshes from foundation models and 2D orientations from specialized predictors trained on limited hair data, restricting efficiency and generalization. We present HairFlow, a multimodal conditioned diffusion framework for efficient and controllable hair strands modeling. HairFlow conditions strand generation on two complementary representations: a coarse mesh that constrains global shape and multi-view sketches that specify local strand flow. HairFlow combines an Orientation VAE with a conditioned diffusion transformer to generate volumetric fields that are traced into explicit strands. To overcome the lack of paired sketch-strand training data, we introduce an automatic synthesis pipeline that generates geometrically aligned multi-view sketches with diverse styles representative of inference-time inputs. Experiments show that HairFlow generates plausible 3D strands that faithfully follow the input conditions across diverse hairstyles, substantially reduces inference time, and supports flexible multimodal control.
Figures & tables
Figure 1: HairFlow uses conditioned diffusion to recover 3D strands from image inputs, guided by 3D mesh and sketches inferred by foundation models. It handles complex hairstyles and stylized images while closely adhering to the input conditions.
Figure 2: Overview of HairFlow. HairFlow takes two geometric inputs—a coarse hair mesh and multi-view sketches—to specify global shape and local strand flow. These conditions can be obtained from images or text using foundation models, or provided by artists. Mesh and camera-aware sketch encoders produce tokens that condition a diffusion transformer through cross-attention. The transformer generates latents in the space of a pretrained Orientation VAE, which decodes them into an orientation field for explicit strand extraction.
Figure 3: Hair sketch synthesis pipeline. (a) Input strands and template head mesh. (b) Multi-view orientation maps; colors encode directed 2D tangents as shown in the circular legend. (c) Bidirectional tracing from sampled seeds; arrows show local directions, and the inset details tracing from the highlighted seed. (d) Sketch variations through stroke density, line width, and localized seed sampling.
Figure 4: Qualitative comparison of single-view 3D hair reconstruction. HairLRM and HairFlow share the same input mesh, supplemented by HairStep orientation maps or sketches from image generation models, respectively. Inputs appear to the left of the corresponding strands. The color wheel below HairLRM illustrates the direction-to-color mapping.
Metric
HairStep
DiffLocks
Im2Haircut
HairLRM
HairFlow (Ours)
IoU ↑
0.625
0.583
0.773
0.855
0.912
Ori-Err ↓
56∘
43∘
35∘
29∘
22∘
Table 1: Quantitative comparison. Silhouette IoU ( ↑ ) is evaluated on 30 real images. Angular error ( ∘ , ↓ ) is measured between 3D orientation fields derived from reconstructed and ground-truth strands on 326 validation hairstyles.
Figure 5: Effect of the sketch condition. From left: input image, mesh, mesh-only result, sketch, and mesh+sketch result. Results within each row share the same mesh and random seed. Colors encode 3D tangent directions.
Method
Training set
Validation set
Mesh Only
Mesh + 2D cues
HairLRM ( Shen et al., 2026 )
10.47
12.70
36.81∘
29.52∘
HairFlow (Ours)
4.01
6.25
27.43∘
22.92∘
Table 2: Additional quantitative comparisons with HairLRM. We report mean angular errors ( ∘ , lower is better) for VAE reconstruction on the training and test sets, and for strand generation conditioned on meshes alone or with additional 2D cues.
Figure 6: Flexible Sketch Guidance: (a) Sketches with different stroke widths and styles yield similar strand geometry. (b) Joint conditioning on sketches from two viewpoints produces a single hairstyle consistent with both views. (c) Two sets of artist-drawn 3D strokes on the same mesh are projected into multi-view sketches, guiding the generated strands to follow different flow patterns.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Qualitative comparisons with HairGPT ( Luo et al., 2026 ) and Hao et al. (2026) . Their input images and reconstruction results are taken from the respective papers, as code is unavailable.
Figure 8: Multi-view renderings of reconstructed strands. Rows 3 and 7 use anime and oil-painting inputs, respectively; row 6 uses multi-view image inputs. The results demonstrate reconstruction across diverse hairstyles, image styles, and input views.
Figure 9: Additional multi-view renderings of reconstructed strands, showing hairstyle geometry and strand flow from different viewpoints.
Figure 10: Additional multi-view results. In the first row, an artist-provided text prompt is used to generate a hair mesh, and sketches are extracted from its rendered images, forming our unified geometric interface for strand generation.
Conditional diffusion models generate or infer data by steering iterative denoising with observations, labels, text, physical constraints, or other inputs.
Hair is a rich medium of visual and cultural expression, yet its digital modeling remains challenging due to the duality of fluidity and structure. Many existing generative approaches rely primarily on continuous diffusion fields, which entangle global topology with local texture and obscure the semantic and structural organization of hairstyles. To address this, we propose HairGPT, a strand-centric framework that treats strands as generative primitives and formulates realistic 3D hairstyle synthesis as a dual-decoupled autoregressive sequence modeling problem. Our method applies spatial decoupling across semantic scalp regions and structural decoupling along a hierarchical strand representation, progressing from global layout to fine-grained style. We further introduce a geometric tokenizer and region-aware semantic annotations to guide strand-level generation, enabling compositional editing, synthesis of rare and complex hairstyles, and adaptation to stylized domains. By aligning generative modeling with the workflow of digital grooming, HairGPT turns hair generation from opaque texture synthesis into a structured and semantically controllable authoring process, supporting robust semantic conditioning and high-fidelity results across realistic and stylized domains. Project Page: https://haiminluo.github.io/hairgpt/
Haimin Luo, Min Ouyang, Lan Xu +1
ShanghaiTech University, China · ShanghaiTech University and Deemos Technology Co., Ltd., China
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.
The fundamental limitation of traditional strand-based modeling is not simply data scarcity, but the ill-posedness of inferring complex 3D fields from 2D imagery without structural constraints. This unconstrained regression leads to catastrophic failures in resolving both global occlusion (e.g., in ponytails) and local directionality (e.g., in curls), resulting in over-smoothed, plausible-but-incorrect geometries. To resolve this, we integrate the strong geometric priors of Large Reconstruction Models (LRMs) into the strand generation pipeline. Using the LRM mesh as a structural anchor, we employ a novel Dual Orientation AutoEncoder to lift coarse geometry into high-fidelity strands. By resolving vector field singularities through latent-space optimization and surface-guided refinement, our method effectively disentangles complex topological structures, setting a new benchmark for robustness and accuracy in hair reconstruction.
Yuefan Shen, Yican Dong, Xiufeng Huang +3
LIGHTSPEED, China · State Key Lab of CAD and CG, Zhejiang University, China · Hong Kong Baptist University, China +1