Pith. sign in

REVIEW 4 major objections 4 minor 50 references

PanoHair: Detailed Hair Strand Synthesis on Volumetric Heads

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PanoHair turns a latent code into a full strand-based hairstyle on a volumetric head in under five seconds for the geometry stages.

desk verdict A genuinely new generative pipeline for hair strands, with a credible fast volume-mesh claim but a circular orientation evaluation that leaves strand quality unsupported. read the letter →

arxiv 2508.18944 v1 pith:HTGRN5JU submitted 2025-08-26 cs.GR cs.CV

Shashikant Verma , Shanmuganathan Raman This is my paper
classification cs.GRcs.CV MSC 68T4568U05
keywords PanoHairstrandsynthesissigneddistancefieldsknowledgedistillation3DgenerativeheadmodelsGabororientationmapslatentcodeinversionvolumeextraction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PanoHair tries to make strand-based hair synthesis generative: instead of multi-view studio images and hours of per-subject optimization, a single latent code should produce a full head geometry, a closed hair volume, and strand geometry. The paper's central claim is that distilling a pre-trained generative head model's tri-grid radiance features into a signed distance field network lets the same latent code drive hair-region semantic labels and view-consistent 3D orientations. If true, hair generation no longer requires complex capture or per-head training: the hair volume and orientation maps are estimated in about five seconds after latent-code inversion, with strand reconstruction following as a separate stage. The authors report higher hair-segmentation IoU and lower projected orientation error than two compared baselines, alongside a large runtime reduction for volume estimation.

What carries the argument

The load-bearing object is the student network Ψ(.), an MLP that decodes tri-grid features from a pre-trained generative head model into signed distance, color, binary hair semantics, and 3D orientation at every sampled point. Around it sit four mechanisms: (1) knowledge distillation, in which the teacher's rendered images and densities supervise the SDF and color; (2) an SDF2Dens conversion that lets the SDF be trained through differentiable volume rendering; (3) a multi-view orientation projection loss that projects predicted 3D orientations into several overlapping views, compares them to Gabor orientation maps, and enforces tangentiality to the surface; and (4) hair-volume closure by fit

What would settle it

Compare PanoHair's grown strands against strand geometry from a CT scan or artist-built model for the same head, measuring angular error of strand directions inside the volume rather than projected surface orientations; the method is falsified if the internal strands point opposite to true growth directions while still matching the projected Gabor maps.

Watch

Extended reading notes

Core claim

On the paper's own terms, PanoHair establishes a generative pipeline whose input is a latent code z and whose outputs are a clean manifold mesh of the hair region, hair-region semantic labels, and view-consistent 3D orientations, all produced in under five seconds once the code is known. The network Ψ(.) is trained by knowledge distillation from a pre-trained generative teacher: the teacher's tri-grid features are decoded into color, density, signed distance, binary hair label, and orientation, with density converted from signed distance via a learned SDF2Dens mapping. The closed hair volume is obtained by fitting a parametric head model, applying a fitted universal scalp placement, and usin

Load-bearing premise

The whole strand-growing stage assumes that Gabor orientation maps computed on synthetic renders are a trustworthy guide to true 3D hair growth direction, even though each Gabor response cannot tell a direction from its opposite and the supervision is never checked against real 3D hair scans.

Editorial extensions

If this is right

  • Generative hair modeling becomes compatible with latent-space operations: interpolating or editing codes should produce smooth changes in hair volume, orientation, and final strands.
  • The five-second geometry estimate makes hair volume and orientation maps cheap enough for interactive or iterative pipelines that previously waited hours.
  • For a real photograph, pivotal inversion into the latent code replaces multi-view acquisition with a single-image setup for novel-view synthesis and strand generation.
  • The hair stages sit entirely on top of the distilled SDF, so improvements to the underlying generative head model should transfer directly to strand quality.
  • The strand-growing stage uses only predicted orientations and the closed hair volume, dropping the silhouette and RGB reconstruction losses used in prior multi-view methods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's evaluation measures projected 2D orientation error rather than agreement of grown strands with true 3D strand geometry; the orientation supervision is never validated against CT scans or artist-built strand models, so a stricter test remains open.
  • Because Gabor responses cannot distinguish θ from θ+π, the learned 3D orientation field may contain globally or locally flipped directions; the multi-view projection loss reduces but does not eliminate this ambiguity, and a strand-direction prior or small amount of 3D annotated data could resolve it.
  • The five-second volume estimate suggests a route to temporally coherent hair editing on video by inverting per-frame latent codes, though the paper does not address temporal stability.
  • The universal FLAME fit is an empirical constant tuned for the bounded geometry of the teacher model; porting the pipeline to another generative head model would likely require re-fitting that constant, exposing a sensitivity point worth testing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. PanoHair is a generative framework that distills a pre-trained PanoHead generator into an SDF network Ψ, which jointly predicts signed distance, color, hair-region semantic labels, and 3D orientations. The orientation field is trained through a multi-view projection loss on Gabor orientation maps rendered from PanoHead (Eq. 7). The hair volume is extracted by fitting FLAME to the generated head via detected landmarks plus a manually chosen 'Universal Fit' translation/scale, followed by boolean subtraction. Strands are then grown inside this volume by aligning strand points to the predicted orientations, following the NeuralHaircut-style pipeline while dropping multi-view image losses. The paper reports superior hair-segmentation IoU, lower Lproj orientation error, and roughly 5-second hair-volume estimation compared with NeuralHaircut and HairStep, along with qualitative strand rendering for synthetic and one inverted real image.

Significance. If the central claims hold, PanoHair would be a practically useful step: it removes the multi-view studio capture requirement for strand-based hair synthesis, provides fast full-head hair-volume extraction, and enables latent-space hairstyle control. The multi-view orientation projection loss is a reasonable attempt to resolve the intrinsic ambiguity of 2D Gabor supervision, and the runtime improvement for volume estimation over NeuralHaircut is notable. The paper ships no code, but the method description is sufficiently precise to reproduce in principle. However, the evaluation does not currently validate the core strand-quality claim. The orientation metric reported in Table 1 is the same Lproj objective used in training, no strand-level geometric metric is provided, and the 'Universal Fit' that defines the hair volume is empirically set without validation. These gaps are addressable but are load-bearing for the paper's main contribution.

major comments (4)
  1. [Table 1; §7.2, Eq. (7)] The orientation error reported in Table 1 is Lproj, the identical loss used to train the orientation head (Eq. 7), and it is measured against Gabor maps computed from the same PanoHead renders used for training. A low value therefore only shows that the network reproduces Gabor-like 2D image structure; it does not measure agreement with true 3D strand tangents. Moreover, Eq. (7) minimizes over IO, IO±π, so the supervision is at most an unoriented line field. The tangential loss (Eq. 6) constrains orientations to the tangent plane but not the growth direction. Since the strand-growing stage (§3.2) is driven entirely by these orientations, this circularity is load-bearing. Please evaluate orientation error against known 3D strand tangents from synthetic hair models with ground-truth strand geometry, or against real 3D hair scans, and report angular error rather than the training loss.
  2. [§4, Figures 6–7; Table 1] The central claim of the paper is 'detailed hair strand synthesis,' yet no strand-level quantitative evaluation is presented. Table 1 measures hair-segmentation IoU and Lproj orientation error only; Figures 6 and 7 show qualitative strand views. There is no comparison of strand geometry (chamfer distance to ground-truth strands, strand count, root positions, strand curvature distribution) against NeuralHaircut, HairStep, or any ground-truth hair model, nor is there a perceptual study. Because the final strand-growing stage optimizes only boundary points against predicted orientations and the predicted volume, without any reconstruction losses, the visual results could partly reflect the smoothness of the optimizer rather than the fidelity of the orientation/volume estimates. A strand-level metric is necessary to support the main contribution.
  3. [§3.2, 'Universal Fit'] The hair volume depends on the 'Universal Fit' of a neutral FLAME model, where the authors state they 'empirically determine translation tf and scale sf' and claim this 'robustly fits all possible generated heads.' No details are provided for how tf and sf are chosen, and no validation is shown across the latent space. This fit defines the scalp boundary, which is the inner surface of the hair volume and the root region for strands, so errors directly propagate to strand roots. Please report the fitting procedure, the chosen values for tf and sf, and quantitative scalp-alignment error over a sample of latent codes, or replace the manual fit with a learned/optimized fit.
  4. [§4, real-image inversion] The abstract and Section 4 claim that for real images the inversion process produces visually appealing hair strands and serves as 'a streamlined alternative to complex multi-view data acquisition.' Figure 7 is the only evidence, and no quantitative evaluation of the inversion is provided. All training and evaluation data are synthetic renders from PanoHead, and the teacher network is PanoHead, so the method's behavior on real photographs is not assessed. If the real-image claim is retained, please add quantitative inversion results (e.g., identity preservation, hair-mask IoU against manual labels, or strand plausibility) on a set of real portraits, or explicitly scope the claims to synthetic generated heads.
minor comments (4)
  1. [Abstract; Table 1] The abstract's 'under 5 seconds' could be misread as the total strand-synthesis time; Table 1 lists strand reconstruction as 4–5 hours. Please clarify that the 5-second claim refers only to hair-volume mesh, semantic maps, and orientation maps.
  2. [Figure 1 caption] The caption says 'distilling knowledge from PanoHair image renderings'; this should presumably read 'PanoHead image renderings.'
  3. [§7.2, Table 2] Table 2 reports columns labeled 'Ltan' and 'Lproj' with values in radians. Please clarify whether these are raw loss values or angular errors, and state the evaluation protocol (e.g., how the 100 sampled latent codes are used and whether the views overlap by 70%).
  4. [Eq. (1)] The piecewise function has mismatched parentheses and brace formatting, and the range or initialization of β is not defined. This is cosmetic but should be fixed for reproducibility.

Circularity Check

2 steps flagged · score 5.0 of 10

Orientation evaluation is the training loss Lproj against the same Gabor maps; semantic IoU reuses the same MODNet training labels; no independent 3D strand validation is provided.

  1. fitted input called prediction [Sec. 3.1 'Loss Functions', Sec. 7.2 (Eq. 7), Table 1]
    "Since ground-truth 3D orientations are unavailable for direct supervision, we instead learn to predict them by enforcing losses on their 2D projections, specifically between O2D and the Gabor orientation map IO. ... Orientation error is measured using Lproj between the projected orientation O2D and high-confidence Gabor orientations."

    The reported 'orientation error' in Table 1 is Lproj, which is exactly the multi-view projection loss minimized during training (Eq. 7). The ground truth for both training and evaluation is the same Gabor orientation map computed from PanoHead renders. A low Lproj therefore only shows that the network reproduces Gabor-filter statistics on the same render distribution used for training; it does not establish that the predicted 3D orientations correspond to true strand tangents. The evaluation metric is the training objective renamed as error, so the orientation-accuracy claim is forced by construction rather than independently tested.

  2. other [Sec. 3.1 'Training Pairs'; Table 1]
    "we utilize [16] to estimate IM, which contains hair and facial masks for I+, and use them as ground truth for supervision. ... We report IoU between the 2D projection of the hair surface and MODNet [16] output as ground truth."

    Semantic segmentation is supervised with BCE(IM, M2D) using MODNet masks IM, and Table 1 evaluates hair-region IoU against those same MODNet masks on renders from the same PanoHead generator. The semantic evaluation therefore reuses the exact label source used for training, demonstrating agreement with MODNet rather than an independently measured hair region. This closes the loop for the semantic sub-claim, though it is less central than the orientation-structure claim.

full rationale

PanoHair's core pipeline is not an equation-level circular derivation: the SDF head geometry is distilled from the pretrained PanoHead teacher, the hair volume is obtained by FLAME fitting and Boolean operations, and strand growth is a post-optimization guided by the predicted field. These components have independent content and produce mesh outputs that can be inspected visually. However, the load-bearing link between the predicted 3D orientation field and the final strand geometry is validated only through Lproj, which is the same loss used to train the network, measured against the same Gabor maps on PanoHead renders. No real 3D strand scans or independent strand-tangent ground truth are used, and the paper itself concedes the directional ambiguity of Gabor responses. The semantic IoU evaluation similarly reuses MODNet masks from the training supervision. These are evaluation-loop circularities: the reported quantitative support for orientation and semantic accuracy reduces to the training objectives, even though the generative mesh and speed claims remain independently meaningful. The undisclosed 'Universal Fit' in Sec. 3.2 is a reproducibility concern but is empirical, not circular. Overall, partial circularity in the quantitative validation, not in the whole derivation, gives a score of 5.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The ledger shows that PanoHair is an assembled pipeline, not a first-principles derivation. Its epistemic weight rests on the PanoHead teacher, on 2D image-space proxies for 3D structure, and on one hand-fitted universal FLAME transform whose values are not disclosed.

free parameters (2)
  • Universal FLAME fit translation t_f and scale s_f = not reported
    Empirically determined in Section 3.2 to position a neutral FLAME scalp for all generated heads. If this fit is wrong, the boolean subtraction produces an incorrect hair volume. It is not validated on real scans.
  • Learnable SDF-to-density sharpness beta = learned, not reported
    The SDF2Dens conversion in Equation 1 uses a learnable beta. This is standard in NeuS-style rendering, but the final learned value is not reported and it directly controls how SDF values are converted to densities.
assumptions (5)
  • domain assumption PanoHead tri-grid features and density renderings provide an accurate full-head geometry and appearance for any latent code z.
    The entire distillation, image reconstruction, and density supervision in Section 3.1 rely on the teacher model being correct. If PanoHead geometry is biased, the student SDF inherits that bias.
  • domain assumption MODNet masks computed on PanoHead renders are valid ground truth for hair semantics.
    The semantic loss in Equation 4 uses BCE between MODNet masks and predicted masks. Errors in MODNet on synthetic renders directly propagate into hair-region segmentation.
  • domain assumption Gabor orientation maps from synthetic renders, combined with a multi-view line-projection loss, are a sufficient proxy for true 3D strand orientation.
    Equation 7 trains orientations against Gabor maps with an explicit pi-ambiguity. The paper notes orientation maps only supervise the outer hair surface, yet no real 3D orientation ground truth is used.
  • domain assumption A single universal FLAME fit can localize the scalp for all generated heads.
    Section 3.2 uses an empirically determined translation and scale to place FLAME inside every generated head. This is a strong generalization claim and is not validated across head shapes or real images.
  • standard math The SDF2Dens mapping in Equation 1 is a valid conversion for volumetric rendering.
    This conversion follows NeuS-style surface rendering and is standard in the literature, but it is an unproved background assumption in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PanoHair: Detailed Hair Strand Synthesis on Volumetric Heads." pith.science (2026). https://pith.science/paper/HTGRN5JU

@misc{pith2026250818944,
  author       = {Pith},
  title        = {Pith review of: PanoHair: Detailed Hair Strand Synthesis on Volumetric Heads},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HTGRN5JU}},
  note         = {Machine review of arXiv:2508.18944}
}
read the original abstract

Achieving realistic hair strand synthesis is essential for creating lifelike digital humans, but producing high-fidelity hair strand geometry remains a significant challenge. Existing methods require a complex setup for data acquisition, involving multi-view images captured in constrained studio environments. Additionally, these methods have longer hair volume estimation and strand synthesis times, which hinder efficiency. We introduce PanoHair, a model that estimates head geometry as signed distance fields using knowledge distillation from a pre-trained generative teacher model for head synthesis. Our approach enables the prediction of semantic segmentation masks and 3D orientations specifically for the hair region of the estimated geometry. Our method is generative and can generate diverse hairstyles with latent space manipulations. For real images, our approach involves an inversion process to infer latent codes and produces visually appealing hair strands, offering a streamlined alternative to complex multi-view data acquisition setups. Given the latent code, PanoHair generates a clean manifold mesh for the hair region in under 5 seconds, along with semantic and orientation maps, marking a significant improvement over existing methods, as demonstrated in our experiments.

Figures

Figures reproduced from arXiv: 2508.18944 by the authors.

Figure 1
Figure 1. PanoHair Overview. Given latent code z and camera parametersCcam = (θ,φ,r), we first obtain tri-grid features Tf using the pre-trained PanoHead [1]. A set of MLPs Ψ(.) then predicts semantic labels, 3D orientations, and SDF values. Training is supervised by distilling knowledge from PanoHair image renderings, Gabor orientations, and semantic seg￾mentation. During inference, only the Blue-boxed components are require… view at source ↗
Figure 2
Figure 2. Inferencing hair’s outer surface and orientations. A 512 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Strand-based hair growth. Given a latent code, (a) shows the generated image, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: Estimating hair volume. (a) FLAME [19] model fitted to estimated landmarks. (b) Empirically computed universal fit, leverag￾ing the structured representation of volumetric faces within a unit volume by design during training. Subsequently, we employ Boolean subtraction…
Figure 5
Figure 5. Figure 5: Latent space interpolation between two face codes and hair growth. The second [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of the proposed approach with state-of-the-art methods, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Latent code optimization using Piv￾otal Inversion [29] to match the given real im￾age (a). The optimized latent code generates novel views (b), hair masks (c), and 3D ori￾entations. The final strand-based hair recon￾struction is shown in (d). Comparisons. While no exis…
Figure 8
Figure 8. Figure 8: The geometric shapes extracted from implicit representations learned using exist [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: For the image (a) generated from a latent code, (b) shows estimated Gabor orienta [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 46 canonical work pages

  1. [1]

    Ogras, and Linjie Luo

    Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y . Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 20950–20959, June 2023

    work page 2023

  2. [2]

    How far are we from solving the 2d & 3d face alignment problem?(and a dataset of 230,000 3d facial landmarks)

    Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem?(and a dataset of 230,000 3d facial landmarks). InProceedings of the IEEE international conference on computer vision, pages 1021–1030, 2017

    2017

  3. [3]

    Single-view hair modeling for portrait manipulation

    Menglei Chai, Lvdi Wang, Yanlin Weng, Yizhou Yu, Baining Guo, and Kun Zhou. Single-view hair modeling for portrait manipulation. ACM Transactions on Graphics (TOG), 31(4):1–8, 2012

    work page 2012

  4. [4]

    Dynamic hair manipulation in images and videos

    Menglei Chai, Lvdi Wang, Yanlin Weng, Xiaogang Jin, and Kun Zhou. Dynamic hair manipulation in images and videos. ACM Transactions on Graphics (TOG), 32(4):1–8, 2013

    work page 2013

  5. [5]

    Efficient geometry-aware 3d generative adversarial networks

    Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16123– 16133, 2022

    work page 2022

  6. [6]

    A statistical wisp model and pseudophysical approaches for interactive hairstyle generation

    Byoungwon Choe and Hyeong-Seok Ko. A statistical wisp model and pseudophysical approaches for interactive hairstyle generation. IEEE Transactions on Visualization and Computer Graphics, 11(2):160–170, 2005. VERMA ET. AL. : PANOHAIR 11

    work page 2005

  7. [7]

    Gram: Generative radiance manifolds for 3d-aware image generation

    Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10673–10683, 2022

    work page 2022

  8. [8]

    Diffusionrig: Learning personalized priors for facial appearance editing

    Zheng Ding, Xuaner Zhang, Zhihao Xia, Lars Jebe, Zhuowen Tu, and Xiuming Zhang. Diffusionrig: Learning personalized priors for facial appearance editing. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12736–12746, 2023

    work page 2023

Show all 50 references
  1. [9]

    Learning neural parametric head models

    Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003–21012, 2023

    2023

  2. [10]

    In search of a general picture processing operator

    Goesta H Granlund. In search of a general picture processing operator. Computer Graphics and Image Processing, 8(2):155–173, 1978

    1978

  3. [11]

    Multiple view geometry in computer vision

    Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision . Cambridge university press, 2003

    2003

  4. [12]

    Sag-free initialization for strand-based hybrid hair simulation

    Jerry Hsu, Tongtong Wang, Zherong Pan, Xifeng Gao, Cem Yuksel, and Kui Wu. Sag-free initialization for strand-based hybrid hair simulation. ACM Transactions on Graphics (TOG), 42(4):1–14, 2023

    2023

  5. [13]

    Collaborative diffusion for multi-modal face generation and editing

    Ziqi Huang, Kelvin CK Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6080–6090, 2023

    2023

  6. [14]

    Nerffaceed- iting: Disentangled face editing in neural radiance fields

    Kaiwen Jiang, Shu-Yu Chen, Feng-Lin Liu, Hongbo Fu, and Lin Gao. Nerffaceed- iting: Disentangled face editing in neural radiance fields. In SIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022

    2022

  7. [15]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119, 2020

    2020

  8. [16]

    Modnet: Real- time trimap-free portrait matting via objective decomposition

    Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real- time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1140–1147, 2022

    2022

  9. [17]

    Diffusionclip: Text-guided diffu- sion models for robust image manipulation

    Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffu- sion models for robust image manipulation. In Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, pages 2426–2435, 2022

    2022

  10. [18]

    Interactive multiresolution hair modeling and editing

    Tae-Yong Kim and Ulrich Neumann. Interactive multiresolution hair modeling and editing. ACM Transactions on Graphics (TOG), 21(3):620–629, 2002

    2002

  11. [19]

    Learning a model of facial shape and expression from 4d scans.ACM Trans

    Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans.ACM Trans. Graph., 36(6):194–1, 2017. 12 VERMA ET. AL. : PANOHAIR

    2017

  12. [20]

    An enhanced framework for real-time hair ani- mation

    Wenqi Liang and Zhiyong Huang. An enhanced framework for real-time hair ani- mation. In 11th Pacific Conference onComputer Graphics and Applications, 2003. Proceedings., pages 467–471. IEEE, 2003

    2003

  13. [21]

    Marching cubes: A high resolution 3d surface construction algorithm

    William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Seminal graphics: pioneering efforts that shaped the field , pages 347–353. 1998

    1998

  14. [22]

    Wide-baseline hair capture using strand-based refinement

    Linjie Luo, Cha Zhang, Zhengyou Zhang, and Szymon Rusinkiewicz. Wide-baseline hair capture using strand-based refinement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 265–272, 2013

    2013

  15. [23]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ra- mamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021

    2021

  16. [24]

    Instant neural graphics primitives with a multiresolution hash encoding

    Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graph- ics (TOG), 41(4):1–15, 2022

    2022

  17. [25]

    Strand-accurate multi-view hair capture

    Giljoo Nam, Chenglei Wu, Min H Kim, and Yaser Sheikh. Strand-accurate multi-view hair capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 155–164, 2019

    2019

  18. [26]

    Modelling and animating cartoon hair with nurbs surfaces

    Paul Noble and Wen Tang. Modelling and animating cartoon hair with nurbs surfaces. In Proceedings Computer Graphics International, 2004., pages 60–67. IEEE, 2004

    2004

  19. [27]

    Capture of hair geometry from multiple images

    Sylvain Paris, Hector M Briceno, and François X Sillion. Capture of hair geometry from multiple images. ACM transactions on graphics (TOG), 23(3):712–719, 2004

    2004

  20. [28]

    Generalized helicoids for modeling hair geometry

    Emmanuel Piuze, Paul G Kry, and Kaleem Siddiqi. Generalized helicoids for modeling hair geometry. InComputer Graphics Forum, volume 30, pages 247–256. Wiley Online Library, 2011

    2011

  21. [29]

    Pivotal tuning for latent-based editing of real images

    Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1): 1–13, 2022

    2022

  22. [30]

    Neural strands: Learning hair geometry and appearance from multi-view images

    Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In European Conference on Computer Vision, pages 73–89. Springer, 2022

    2022

  23. [31]

    3d hair synthesis using volumetric variational autoencoders

    Shunsuke Saito, Liwen Hu, Chongyang Ma, Hikaru Ibayashi, Linjie Luo, and Hao Li. 3d hair synthesis using volumetric variational autoencoders. ACM Transactions on Graphics (TOG), 37(6):1–12, 2018

    2018

  24. [32]

    Ct2hair: High-fidelity 3d hair modeling using computed tomography

    Yuefan Shen, Shunsuke Saito, Ziyan Wang, Olivier Maury, Chenglei Wu, Jessica Hod- gins, Youyi Zheng, and Giljoo Nam. Ct2hair: High-fidelity 3d hair modeling using computed tomography. ACM Transactions on Graphics (TOG), 42(4):1–13, 2023. VERMA ET. AL. : PANOHAIR 13

    2023

  25. [33]

    Neural haircut: Prior-guided strand-based hair reconstruction

    Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempit- sky, and Egor Zakharov. Neural haircut: Prior-guided strand-based hair reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19762–19773, 2023

    2023

  26. [34]

    Text-conditioned generative model of 3d strand-based human hairstyles

    Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J Black, and Justus Thies. Text-conditioned generative model of 3d strand-based human hairstyles. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4703–4712, 2024

    2024

  27. [35]

    Semfaceedit: Semantic face editing on generative radiance manifolds

    Shashikant Verma and Shanmuganathan Raman. Semfaceedit: Semantic face editing on generative radiance manifolds. In International Conference on Pattern Recognition, pages 46–62. Springer, 2025

    2025

  28. [36]

    Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021

    2021 arXiv

  29. [37]

    Hair design based on the hierarchical cluster hair model

    Tao Wang and Xue Dong Yang. Hair design based on the hierarchical cluster hair model. Geometric modeling: techniques, applications, systems and tools , pages 329– 359, 2004

    2004

  30. [38]

    A trigonal prism-based method for hair image generation

    Yasuhiko Watanabe and Yasuhito Suenaga. A trigonal prism-based method for hair image generation. IEEE Computer Graphics and applications, 12(01):47–53, 1992

    1992

  31. [39]

    Neural- hdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations

    Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neural- hdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages ...

    2022

  32. [40]

    Monohair: High-fidelity hair modeling from a monocular video

    Keyu Wu, Lingchen Yang, Zhiyi Kuang, Yao Feng, Xutao Han, Yuefan Shen, Hongbo Fu, Kun Zhou, and Youyi Zheng. Monohair: High-fidelity hair modeling from a monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24164–24173, 2024

    2024

  33. [41]

    Dynamic hair modeling from monocular videos using deep neural networks

    Lingchen Yang, Zefeng Shi, Youyi Zheng, and Kun Zhou. Dynamic hair modeling from monocular videos using deep neural networks. ACM Transactions on Graphics (TOG), 38(6):1–12, 2019

    2019

  34. [42]

    Multiview neural surface reconstruction by disentangling geometry and appearance

    Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492–2502, 2020

    2020

  35. [43]

    Human hair reconstruction with strand-aligned 3d gaussians

    Egor Zakharov, Vanessa Sklyarova, Michael Black, Giljoo Nam, Justus Thies, and Ot- mar Hilliges. Human hair reconstruction with strand-aligned 3d gaussians. InEuropean Conference on Computer Vision, pages 409–425. Springer, 2024

    2024

  36. [44]

    Hair-gan: Recovering 3d hair structure from a single image using generative adversarial networks

    Meng Zhang and Youyi Zheng. Hair-gan: Recovering 3d hair structure from a single image using generative adversarial networks. Visual Informatics, 3(2):102–112, 2019. 14 VERMA ET. AL. : PANOHAIR

    2019

  37. [45]

    Hairstep: Transfer synthetic to real using strand and depth maps for single-view 3d hair modeling

    Yujian Zheng, Zirong Jin, Moran Li, Haibin Huang, Chongyang Ma, Shuguang Cui, and Xiaoguang Han. Hairstep: Transfer synthetic to real using strand and depth maps for single-view 3d hair modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2023

  38. [46]

    Hairnet: Single-view hair reconstruction using convolutional neural networks

    Yi Zhou, Liwen Hu, Jun Xing, Weikai Chen, Han-Wei Kung, Xin Tong, and Hao Li. Hairnet: Single-view hair reconstruction using convolutional neural networks. In Pro- ceedings of the European Conference on Computer Vision (ECCV) , pages 235–251, 2018

    2018

  39. [47]

    Sean: Image synthesis with semantic region-adaptive normalization

    Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5104–5113, 2020

    2020

  40. [48]

    Instant volumetric head avatars

    Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 4574–4584, 2023. VERMA ET. AL. : PANOHAIR 15 Appendix 6 PanoHead Why PanoHead Over Existing Approac...

    2023

  41. [49]

    PanoHead [1] further advances this by leveraging Tri-Grid feature representations

    employs Tri-Plane representations. PanoHead [1] further advances this by leveraging Tri-Grid feature representations. As illustrated in Figure 8, methods trained solely on frontal views struggle to accurately capture the complete 360◦ geometry of the human head, posing a chall...

  42. [50]

    In contrast, we leverage knowledge distillation from PanoHead to learn full-head SDFs (Figure 8(d)), whereas PanoHead represents faces as density distributions (Figure 8(c))

    and Tri-planes [5] primarily reconstruct frontal geometries (Figure 8(a), (b)). In contrast, we leverage knowledge distillation from PanoHead to learn full-head SDFs (Figure 8(d)), whereas PanoHead represents faces as density distributions (Figure 8(c)). 7 Loss Functions In th...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.