SAFE-Hair: Scalp-Anchored Fields for Exportable Single-View 3D Hair Reconstruction

Contributions

32,768 strands

Grown from fixed follicles on a shared scalp, from one portrait.
Persistent identity

Every strand can be edited, simulated and exported in one forward pass.
Scalp-rooted

The structure grooming and simulation need. Meshes and Gaussians lose it.

Mesh-based

No explicit strands, so nothing can be groomed or simulated per strand, and there is no follicle correspondence.
Gaussian-based

Strands pass through the head, follicles are unstable and do not persist, and the result is hard to serialise as a groom.
SAFE-Hair (ours)

Explicit strands rooted on the scalp, a persistent follicle ID for every one by construction, and collision-aware grooms ready to export.

Methodology

Instead of predicting unordered strands in image or world space, SAFE-Hair reasons entirely in a canonical scalp-UV domain shared by all subjects, then decodes one candidate strand per persistent follicle.

SAFE-Hair pipeline overview — Figure 2. **Overview.** Dense image features are projected into scalp-UV space; a conditional Rectified Flow turns noise into a latent scalp field \(Z\); a shared follicle decoder queries \(Z\) at canonical follicles \(\{\mathbf{u}_r\}_{r=1}^{R}\) and strands are obtained by forward integration from fixed scalp roots.

Scalp-UV conditioning

Fit a template head, extract DINOv2 features and warp them onto the scalp through the recovered camera. A fixed follicle set gives every strand a persistent identity.

\[ \mathcal{R}=\{\mathbf{u}_r\}_{r=1}^{R},\qquad \mathbf{x}_{r,1}=\psi(\mathbf{u}_r) \]
Rectified flow on scalp fields

A conditional vector field transports Gaussian noise to a latent scalp field \(Z\in\mathbb{R}^{128\times128\times64}\), completing invisible regions with learned priors.

\[ \mathcal{L}_{\mathrm{rf}}=\mathbb{E}\big\|v_\theta(z_t,t,C)-(z_1-z_0)\big\|_2^2 \]
Persistent follicle decoding

A shared decoder predicts occupancy, segment directions, length allocation and total length; strands grow from the root by forward integration.

\[ \mathbf{x}_{r,t+1}=\mathbf{x}_{r,t}+L_r\,q_{r,t}\,\hat{\mathbf{d}}_{r,t} \]
Validity priors and export

Attachment, collision and bending priors reduce root flipping, head interpenetration and kinks. Strands are grouped into guide families and serialised to Alembic or engine-native grooms.

\[ g(r)=\arg\min_m\ \alpha\|\mathbf{u}_r-\mathbf{u}_m\|^2+\beta\|\mathbf{f}_r-\mathbf{f}_m\|^2 \]

Interactive: grow strands from fixed roots

Each curve is integrated from a fixed root on the scalp using per-segment directions \(\hat{\mathbf{d}}_{r,t}\) and length weights \(q_{r,t}\). Switch the priors off to see why they matter.

Total length \(L_r\) Curl

Collision prior Attachment prior Colour by guide family

Penetrating samples: 0%

Qualitative Results

The portrait shows only the front. SAFE-Hair completes the side and back on the scalp, and curls continue around the head instead of stopping at the silhouette.

Input one front-facing portrait
Front reconstructed groom
Side never seen in the input
Back completed by learned priors

Comparison with Img2Haircut and DiffLocks

Against Img2Haircut and DiffLocks, SAFE-Hair produces scalp-attached strand geometry with more consistent back-view completion and fewer head-interpenetration artifacts.

Quantitative Results

On HairBench-3D, SAFE-Hair leads every geometry metric and every asset metric against five baselines. It reaches an F-score of 81.7 (from 77.9), 69.8 on invisible regions (from 66.1) and a penetration ratio of 3.9% (from 4.8%), compared with the strongest prior baseline.

99.8%

of roots land on a valid scalp region, against 93.2% for the strongest prior method.
3.28 mm

Chamfer distance, down from 3.74 mm.
4.2 s

per portrait for 32,768 strands, with no test-time optimisation.

Main results

Method	ChamferPts ↓	F-score ↑	F-score (Invis.) ↑	IoU ↑	Pen. ratio ↓
Perm	5.82	61.4	42.7	0.781	8.6
NeuralHDHair	4.36	71.8	56.2	0.824	6.1
Hairstep	4.48	70.6	59.4	0.872	7.2
DiffLocks	3.91	76.5	63.8	0.847	5.4
Im2Haircut	3.74	77.9	66.1	0.861	4.8
SAFE-Hair (ours)	3.28	81.7	69.8	0.884	3.9

Asset usability

Method	RootDist ↓	ValidRoot ↑	FollicleStab. ↑	ExportSucc. ↑
Perm	–	–	–	–
NeuralHDHair	3.8	82.1	–	71.3
Hairstep	2.9	87.4	–	76.8
DiffLocks	2.4	89.6	0.61^†	82.5
Im2Haircut	1.7	93.2	0.68^†	88.1
SAFE-Hair (ours)	0.2	99.8	0.91	99.6

Bold: best; underlined: second best. RootDist: mean root-to-scalp distance (mm). ValidRoot: % of roots on a valid scalp region. FollicleStab.: active-follicle consistency under crop, colour and mask perturbations. ExportSucc.: % of predictions serialisable as valid grooms. ^†Approximated by nearest-root matching.

Exportable Groom

Every strand is decoded from a canonical follicle, so its identity is simply the follicle index \(r\). Follicles are grouped into guide families and the result serialises straight into Alembic or engine-native strand assets.

(a) Guide-family clusters
(b) Guide strands
(c) Refined groom

Dense follicles. 32,768 canonical follicles with 32 control points per strand.
Guide selection. Guides chosen by farthest-point sampling in scalp-UV space.
Child assignment. Children assigned in joint scalp-position and strand-feature space.
Lossless metadata. Grouping never alters the reconstructed geometry.

strand {
  follicle_id: 18342
  scalp_uv: (0.412, 0.637)
  points: float3[32]
  guide_id: 207
}

BibTeX

@inproceedings{wang2026safehair,
  title     = {SAFE-Hair: Scalp-Anchored Fields for Exportable
               Single-View 3D Hair Reconstruction},
  author    = {Wang, Yucheng and Wang, Zedong and Wu, Yuetong
               and Ma, Yue and Xu, Dan},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}

Contributions

32,768 strands

Persistent identity

Scalp-rooted

Mesh-based

Gaussian-based

SAFE-Hair (ours)

Methodology

Scalp-UV conditioning

Rectified flow on scalp fields

Persistent follicle decoding

Validity priors and export