SAFE-Hair: Scalp-Anchored Fields for Exportable Single-View Hair Reconstruction

Yucheng Wang; Zedong Wang; Yuetong Wu; Yue Ma; Dan Xu

cs.CVNov 9, 2026

SAFE-Hair: Scalp-Anchored Fields for Exportable Single-View Hair Reconstruction

Authors: Yucheng Wang, Zedong Wang, Yuetong Wu, Yue Ma, Dan Xu

Organizations: The Hong Kong University of Science and Technology

Abstract

Single-view 3D hair reconstruction has progressed rapidly, but many methods optimize for visual plausibility rather than producing strand assets that can be directly used in grooming, editing, or simulation pipelines. We propose SAFE-Hair, a framework that reconstructs explicit strand grooms from a single portrait by decoding hair from persistent follicles in a canonical scalp-UV domain. Given a fitted template head, SAFE-Hair projects self-supervised image features onto the scalp, uses a conditional Rectified Flow model to generate latent scalp fields, and decodes occupancy, segment directions, length allocation, and total length at a fixed set of scalp follicles. This representation enforces scalp-rooted strand identities by construction and enables direct export with follicle IDs and guide-family metadata. We further introduce attachment, collision, and bending priors to improve geometric validity without per-subject test-time optimization. To evaluate both reconstruction fidelity and asset usability, we introduce HairBench-3D, a benchmark built from public 3D hair assets with standardized alignment, rendering, resampling, and metrics for geometry, direction consistency, invisible-region completion, silhouette agreement, and collision. SAFE-Hair improves Chamfer distance from 3.74 mm to 3.28 mm, F-score from 77.9 to 81.7, invisible-region F-score from 66.1 to 69.8, and penetration ratio from 4.8% to 3.9% over the strongest baseline.

Figures & tables

Figure 1: From a single portrait image, SAFE-Hair reconstructs an explicit strand-based 3D hair asset with stable scalp-rooted follicles. We visualize each reconstruction from front, side, and back views, showing complete hairstyle recovery beyond the visible input region. The bottom row highlights the diversity of generated assets across multiple hairstyle categories.

Figure 2: Motivation. Existing 3D hair generators often produce rendering-oriented representations: Mesh-based hair [ 39 ] lacks explicit strands, while Gaussian-based hair (UniHair [ 41 ] ) lacks persistent follicles and suffers from collisions. Our method reconstructs explicit rooted strands with stable scalp follicles, enabling editable and export-ready hair assets from a single portrait.

Figure 3: Overview of SAFE-Hair. Given a single portrait, we fit a template head and use the recovered camera geometry to project dense image features into a canonical scalp-UV domain, producing a scalp-aligned conditioning map. A conditional Rectified Flow model transforms Gaussian noise into a latent scalp field $Z$ . A shared follicle decoder queries $Z$ at persistent canonical follicles $\{\mathbf{u}_{r}\}_{r=1}^{R}$ and predicts occupancy, segment directions, length allocation, total length, and export descriptors. The final strand curves are obtained by forward integration from fixed scalp roots, yielding explicit scalp-rooted strands with stable follicle identities and groom metadata for downstream export.

Source	#Assets	Main role
Hair20K [ 43 ]	1,200	Diverse hairstyles
USC-HairSalon [ 10 ]	343	Groom assets
Yuksel [ 34 ]	8	Varied examples
Total	1,551	Frozen test set

Table 1: HairBench-3D held-out test split. The split is defined at the hairstyle-asset level.

Figure 4: Qualitative comparison. For each input image, we show reconstructed hair from multiple viewpoints. SAFE-Hair produces scalp-attached strand geometry with more consistent back-view completion and fewer head-interpenetration artifacts than prior methods.

Method	ChamferPts $\downarrow$	F-score $\uparrow$	F-score (Invisible) $\uparrow$	IoU $\uparrow$	Penetration Ratio $\downarrow$
Perm [ 7 ]	5.82	61.4	42.7	0.781	8.6
UniHair [ 41 ]	4.85	69.2	57.0	0.807	26.8
NeuralHDHair [ 32 ]	4.36	71.8	56.2	0.824	6.1
Hairstep [ 40 ]	4.48	70.6	59.4	0.872	7.2
DiffLocks [ 26 ]	3.91	76.5	63.8	0.847	5.4
Im2Haircut [ 28 ]	3.74	77.9	66.1	0.861	4.8

Table 2: Main results on HairBench-3D. Lower is better for ChamferPts and Penetration Ratio, while higher is better for F-score@4mm/40 ∘ , Invisible-region F-score, and Silhouette IoU. “Ours” uses one fixed seed; “Ours (5 seeds)” reports mean

\pm

standard deviation over five Rectified Flow sampling seeds with fixed weights.

Figure 5: Qualitative ablation. Left: SAFE-Hair produces more complete and coherent scalp-rooted strands. Right: removing scalp-UV conditioning or the latent scalp field degrades strand coherence and hairstyle consistency.

Variant	ChamferPts $\downarrow$	F-score $\uparrow$	F-score (Invisible) $\uparrow$	IoU $\uparrow$	Penetration Ratio $\downarrow$
W/o Scalp UV	3.91	77.8	63.5	0.857	5.1
W/o Rectified Flow	3.67	78.9	64.1	0.866	4.8
W/o Latent Scalp Field	3.58	79.6	66.0	0.871	4.6
W/o Length Factorization	3.49	80.2	67.1	0.876	4.4
W/o Directions	3.43	80.8	68.2	0.879	6.7
Full SAFE-Hair	3.28	81.7	69.8	0.884	3.9

Table 3: Ablation study on HairBench-3D. We evaluate the contribution of the key design choices in SAFE-Hair using the primary benchmark metrics, including geometry accuracy, visible and invisible-region reconstruction quality, silhouette consistency, and physical validity.

Method	RootDist $\downarrow$	ValidRoot $\uparrow$	FollicleStab. $\uparrow$	ExportSucc. $\uparrow$
Perm [ 7 ]	–	–	–	–
NeuralHDHair [ 32 ]	3.8	82.1	–	71.3
Hairstep [ 40 ]	2.9	87.4	–	76.8
DiffLocks [ 26 ]	2.4	89.6	0.61 †	82.5
Im2Haircut [ 28 ]	1.7	93.2	0.68 †	88.1
Ours	0.2	99.8	0.91	99.6

Table 4: Asset usability on HairBench-3D. RootDist measures average root-to-scalp distance in mm; ValidRoot and ExportSucc. are percentages. “–” marks metrics that are not applicable, and † indicates approximate follicle stability computed by nearest-root matching. Metric definitions are given in Appendix B.6 .

Input yaw	ChamferPts $\downarrow$	F-score $\uparrow$	F-score (Invisible) $\uparrow$	IoU $\uparrow$
$0^{\circ}$	3.28	81.7	69.8	0.884
$30^{\circ}$	3.35	81.0	68.9	0.878
$60^{\circ}$	3.57	79.4	66.2	0.861
$90^{\circ}$	3.86	77.2	62.8	0.839

Table 5: Generalization to side-view inputs on HairBench-3D. Each held-out hairstyle is rendered at four input yaw angles and reconstructed from the corresponding portrait.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Steps	Time (s) $\downarrow$	ChamferPts $\downarrow$	F-score $\uparrow$	F-score (Invisible) $\uparrow$
5	1.4	3.47	79.8	67.4
10	2.4	3.34	81.0	69.0
20	4.2	3.28	81.7	69.8
40	7.8	3.26	81.9	70.0

Appendix

Table 6: Effect of Rectified Flow sampling steps on HairBench-3D.

Follicles $N$	Points $T$	Time (s) $\downarrow$	Memory (GB) $\downarrow$	ChamferPts $\downarrow$	F-score $\uparrow$
32,768	32	4.2	7.9	3.28	81.7
65,536	32	6.8	12.7	3.22	82.1
32,768	64	5.6	10.8	3.21	82.0
65,536	64	9.4	18.6	3.17	82.5

Appendix

Table 7: Scalability in follicle count

N

and control points per strand

T

Panel of Figure 6 — Figure 6: Exportable groom structure. SAFE-Hair groups active follicles into guide families in canonical scalp UV space. Each child strand stores a persistent follicle ID, guide ID, control points, and optional physical or groom-control attributes.

Usage of HairBench-3D test split	Allowed?
Training data	No
Validation or hyperparameter tuning	No
Loss-weight selection	No
Filtering-threshold selection	No
Qualitative example selection before final evaluation	No
Final quantitative evaluation	Yes

Appendix

Table 8: Usage policy for the frozen HairBench-3D test split.

Perturbation	ChamferPts $\downarrow$	F-score $\uparrow$	RootDist-to-GT $\downarrow$	ValidRoot $\uparrow$
None	3.28	81.7	0.20	99.8
Camera rotation ( $\pm 5^{\circ}$ )	3.33	81.2	0.22	99.6
Head-shape deviation	3.53	79.5	0.96	97.8
Combined	3.62	78.9	1.15	96.7

Appendix

Table 9: SAFE-Hair under head-and-camera fitting errors.

	Camera rotation				Head-shape deviation				Combined
Method	$\Delta$ CD	$\Delta$ F	$\Delta$ Root	$\Delta$ Valid	$\Delta$ CD	$\Delta$ F	$\Delta$ Root	$\Delta$ Valid	$\Delta$ CD	$\Delta$ F	$\Delta$ Root	$\Delta$ Valid
DiffLocks [ 26 ]	2.2%	0.8	0.08	0.8	9.8%	3.1	1.25	4.8	14.2%	4.1	1.52	6.1
Im2Haircut [ 28 ]	2.0%	0.7	0.05	0.5	9.4%	2.8	1.02	3.8	12.6%	3.7	1.24	5.0
Ours	1.5%	0.5	0.02	0.2	7.6%	2.2	0.76	2.0	10.4%	2.8	0.95	3.1

Appendix

Table 10: Robustness to head-and-camera fitting errors. Degradation from each method’s clean setting (lower is better);

\Delta

F and

\Delta

Valid in points,

\Delta

Root in mm.

Figure 7: Limitation. Extreme hairstyles.

Figure 8: Additional qualitative comparison on in-the-wild portrait images. From left to right: input image, SAFE-Hair, Hairstep, and NeuralHDHair. SAFE-Hair produces more coherent and complete strand reconstructions across diverse hairstyles, including short, straight, wavy, spiky, curly, and afro hair. Compared with baselines, our method better preserves the global hair silhouette and local strand flow, while reducing floating strands, sparse regions, and unstable scalp attachment.

Figure 9: Additional qualitative results on in-the-wild portrait images. For each pair, the left image shows the input portrait and the right image shows the reconstructed 3D hair rendered on a neutral template head. These examples show the diversity of single-image reconstructions produced by SAFE-Hair.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 992 papers

HairOrbit: Multi-view Aware 3D Hair Modeling from Single Portraits
Apr 3, 2026Leyang Jin, Yujian Zheng, Bingkui Tong +33D Reconstruction Single-View 3D Reconstruction
Mohamed bin Zayed University of Artificial Intelligence · SSE, The Chinese University of Hong Kong, Shenzhen · Pinscreen
Wispy to Voluminous: Prior-free Multi-view Capture of Strand-level Facial Hair
Jun 6, 2026Jaeseong Lee, Giljoo Nam, Adrian Jarabo +13D Reconstruction Multi-View 3D Reconstruction
1KAIST · 3Meta Reality Labs Research · 2Meta Codec Avatar Lab
HairLRM: Strand-based Hair Modeling via Large Reconstruction Models
Jun 13, 2026Yuefan Shen, Yican Dong, Xiufeng Huang +33D Reconstruction Generative Modeling
LIGHTSPEED, China · State Key Lab of CAD and CG, Zhejiang University, China · Hong Kong Baptist University, China +1

Apr 3, 2026J/K move · Enter open · S save

HairOrbit: Multi-view Aware 3D Hair Modeling from Single Portraits

Leyang Jin, Yujian Zheng, Bingkui Tong, Yuda Qiu+2

Mohamed bin Zayed University of Artificial Intelligence · SSE, The Chinese University of Hong Kong, Shenzhen · Pinscreen

3D Reconstruction Single-View 3D Reconstruction

Open paper AskPDF arXiv

Reconstructing strand-level 3D hair from a single-view image is highly challenging, especially when preserving consistent and realistic attributes in unseen regions. Existing methods rely on limited frontal-view cues and small-scale/style-restricted synthetic data, often failing to produce satisfactory results in invisible regions. In this work, we propose a novel framework that leverages the strong 3D priors of video generation models to transform single-view hair reconstruction into a calibrated multi-view reconstruction task. To balance reconstruction quality and efficiency for the reformulated multi-view task, we further introduce a neural orientation extractor trained on sparse real-image annotations for better full-view orientation estimation. In addition, we design a two-stage strand-growing algorithm based on a hybrid implicit field to synthesize the 3D strand curves with fine-grained details at a relatively fast speed. Extensive experiments demonstrate that our method achieves state-of-the-art performance on single-view 3D hair strand reconstruction on a diverse range of hair portraits in both visible and invisible regions.

Abstract

3D Reconstruction

Explore similar work

HairOrbit: Multi-view Aware 3D Hair Modeling from Single Portraits

Wispy to Voluminous: Prior-free Multi-view Capture of Strand-level Facial Hair

HairLRM: Strand-based Hair Modeling via Large Reconstruction Models