Pith. sign in

REVIEW 4 major objections 7 minor 19 references

HairFormer: Transformer-Based Dynamic Neural Hair Simulation

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Two-stage Transformer network predicts static and dynamic hair for unseen hairstyles, body shapes, and motions in real time.

desk verdict A genuine two-stage transformer architecture for real-time hair draping and dynamics, but the printed functional-regularization loss is vacuous and the quantitative evaluation partly re-measures the training objectives. read the letter →

arxiv 2507.12600 v1 pith:4BBNVXBT submitted 2025-07-16 cs.GR cs.CV

Joy Xiaoji Zhang , Jingsen Zhu , Hanyu Chen , Steve Marschner This is my paper · ORCID
classification cs.GRcs.CV
keywords neuralhairsimulationreal-timedynamicsTransformernetworkscross-attentionfusionphysics-informedlossesself-supervisedlearninghair-bodypenetrationresolutionstrand-basedrepresentation
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

This paper claims that a single two-stage neural network can simulate hair dynamics that generalize to arbitrary hairstyles, body shapes, and motions, with no per-asset simulation. The static stage uses a Transformer with hair-to-body cross-attention to predict a draped shape that resolves hair-body penetrations; the dynamic stage fuses those static features with motion descriptors through a recurrent cross-attention mechanism to produce secondary motions such as flying hair. The authors report real-time inference for both stages and argue that this is the first method to combine such broad generalization with expressive dynamics. The contribution matters because realistic hair is one of the hardest remaining bottlenecks for real-time digital humans, and existing learned methods are mostly quasi-static or tied to particular hairstyles.

What carries the argument

The load-bearing representation is a coarse grid of deformation maps on the scalp (for example 8×8 cells), where each hair strand looks up its deformation by root UV coordinate and a tapering function scales deformations down toward the root for smooth attachment. The static network is a Transformer encoder: hair latent tokens are processed with self-attention and then modulated by a body context vector through cross-attention, and the output tokens are reshaped into the deformation map. The dynamic network passes motion descriptors (first-order pose derivatives and joint accelerations over a window of frames) through a GRU, then applies static-dynamic cross-attention with the static hair features as queries and the dynamic latent as keys and values; a functional-regularization loss keeps the dynamic prediction head consistent with the frozen static head. Physics is imposed only through weighted loss energies: inextensibility, bending, shape auxiliary and smoothness, gravity, contact barriers for hair-body and hair-hair, root alignment, and inertia. These losses are the sole physics in the system and are also the metrics used to report quantitative results.

What would settle it

Run HairFormer and a high-fidelity physics reference such as Discrete Elastic Rods on the same long hairstyle and a fast head-rotation sequence, then compare per-strand trajectories over time; if the predicted motion diverges strongly from the reference on in-distribution poses, or if hair-body penetration persists around complex regions such as the chest and unmodeled hand joints on unseen body shapes, the paper's claims of physical plausibility and broad generalization would be undercut.

Watch

Extended reading notes

Core claim

The central claim is that physically plausible hair dynamics for arbitrary and unseen hairstyles can be learned directly from physics-informed losses, without any pre-computed simulation data or reference simulator. Hair strands are generated from a StyleGAN-based hair latent code, rigidly posed onto an SMPL body, and then a Transformer-based static network predicts a per-vertex deformation map that drapes the hair and pushes it out of the body using an extended C-IPC-style barrier function, even for negative signed distances. A dynamic network then takes windowed motion descriptors through a GRU and uses a cross-attention mechanism to combine static hair features with dynamic context, producing per-timestep deformations that preserve hairstyle identity while adding complex secondary motion. The paper reports that this design resolves penetrations for complex unseen long hairstyles, preserves hair length better than an adapted Quaffure baseline and direct optimization, and can be fine-tuned in a few hundred iterations to handle abrupt head movements such as whiplash.

Load-bearing premise

The hand-weighted sum of physics-inspired losses is assumed to be a faithful and sufficient model of real hair physics, with no reference simulator or measured data to calibrate the weights.

Editorial extensions

If this is right

  • A single trained model can drape and animate hairstyles never seen during training, removing per-asset simulation for digital humans.
  • Static and dynamic inference both run in real time, with roughly 4.6 ms per static drape and 12.2 ms per dynamic drape on the reported hardware.
  • The dynamic network produces secondary motions such as flying and antenna hair on unseen pose sequences without any fine-tuning.
  • Challenging sequences like whiplash head motions can be adapted in a few hundred iterations of fine-tuning.
  • The extended barrier function resolves hair-body penetrations better than the quasi-static baseline and direct optimization while preserving hair length in dynamic sequences.

Reading between the lines

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

  • Beyond the paper, the same cross-attention recipe could be retrained with a calibrated reference simulator or real hair capture replacing the hand-set loss weights, which would decouple fidelity from manual tuning.
  • If the architecture is as style-agnostic as claimed, it should transfer to other strand-like deformables such as fur, yarn, or ropes, where a latent code plus body kinematics plays the role of hair style and SMPL pose.
  • The dynamic network's reliance on pose-derived velocity and acceleration means that changing playback speed currently demands retraining or fine-tuning; exposing velocity as an explicit input would enable retimed animations without retraining.
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 / 7 minor

Summary. The paper introduces HairFormer, a two-stage Transformer-based neural hair simulation method. A static network takes a StyleGAN hair latent, SMPL body shape, and pose to predict per-vertex deformation offsets that drape the hair over the body, trained with self-supervised physics-inspired losses including an IPC-style barrier for hair-body contact. A dynamic network then fuses static hair features with recurrent motion descriptors via cross-attention to generate temporally coherent hair motion, also trained with an inertia loss and a functional regularization term. The authors claim real-time inference, generalization to unseen hairstyles and body motions, and effective penetration resolution, and they report quantitative comparisons to adapted Quaffure and direct optimization baselines.

Significance. If the claims are substantiated, this work would be a meaningful step toward real-time, generalizable hair simulation that avoids per-asset simulation and does not require pre-simulated training data. The self-supervised formulation, the use of Transformer cross-attention for hair-body and static-dynamic feature fusion, and the explicit barrier-based penetration handling are well-aligned with current research directions. The paper provides qualitative results on unseen hairstyles, an ablation study of the loss components and fusion mechanisms, and reports inference times on the order of milliseconds. However, the significance is tempered by evaluation that largely re-measures the training objectives and by a load-bearing equation error that leaves the functional regularization mechanism undefined as printed.

major comments (4)
  1. [Sec. 3.3.4, Eq. (4)] The functional regularization loss as printed is identically zero: L_func_reg = lambda_fr * MSE(F_dyn_pred(H_static), F_dyn_pred(H_static)) compares the dynamic prediction head with itself on the static features. The surrounding text states that the goal is to keep the dynamic head's behavior similar to that of the frozen static head, so the second argument should be F_static_pred(H_static) or the equation should be corrected otherwise. As written, this term provides no gradient, so the claimed mechanism for preserving static features during dynamic training is not present in the equations, and the ablation in Sec. 4.5.2 that relies on this loss is not reproducible from the paper.
  2. [Sec. 4.3, Table 1] The quantitative evaluation is partly circular: the two reported metrics, Penetration (%) and Length Change (%), are direct reflections of the contact loss L_hb and the inextensibility loss L_s used for training. Reporting these metrics shows that the network minimizes its own training objectives, but it does not demonstrate physical fidelity relative to an external reference. The paper does not compare against a standard physics-based simulator or measured hair data, nor does it provide error bars or multiple test scenes. I recommend adding an external ground-truth comparison (e.g., a well-validated simulation reference) and evaluating on several varied scenes with statistical significance.
  3. [Abstract and Limitations (Sec. 5)] The abstract's claim of 'effectively resolving penetrations' is not supported by the quantitative results in Table 1, which show 11.461% penetration for the static result and 6.751% for the dynamic result on the single reported test scene. The Limitations section also explicitly acknowledges persisting penetrations around the breast area and unmodeled hand joints, and Fig. 8 shows residual left-shoulder penetration for the long-hair case. The claims should be tempered or complemented by a discussion of why these penetration levels are acceptable in practice, and by evidence that the method generalizes beyond the single scene shown in Table 1.
  4. [Secs. 3.4 and 4.2] The paper omits essential values for the physics-based losses and training procedure. Loss weights k_s, k_b, lambda_aux, lambda_smooth, lambda_hb, lambda_hh, lambda_root, lambda_inertia, and lambda_fr are never reported; similarly, vertex masses m_s,v', the hard/soft barrier distances xi and d_hat, the deep-penetration distance t0, and the dynamic-score filtering threshold are unspecified. Since the method is self-supervised and the physical plausibility of the output depends directly on these manually chosen weights, their absence prevents reproduction and makes it impossible to assess sensitivity of the results to these choices. A complete hyperparameter table should be added.
minor comments (7)
  1. [Eq. (12)] The piecewise barrier function definition is hard to parse because the denominator and exponents are not clearly typeset; please define xi and d_hat before the equation and use a cleaner notation for the squared terms.
  2. [Sec. 3.3.4] The sentence 'we seek to encourage H_dyn to evolve within a compatible latent space' is vague; please clarify precisely how enforcing consistency of the prediction heads achieves compatibility of the latent features.
  3. [Fig. 2 caption] The caption repeats much of the main text; it could be shortened to improve readability.
  4. [References] The reference 'Vaswani 2017' is incomplete (missing all authors and full title); the Li et al. 2020 reference appears to be the C-IPC paper, but the citation in Sec. 3.4.4 should be explicit about the codimensional extension.
  5. [Table 1] The time units are given as 'ms/drape' in the header but the values are labeled 'ms' in the text; also missing are the number of test samples and any error bars.
  6. [Sec. 3.1] The notation M_hair (number of strands) is close to M_H0 x M_W0 (deformation grid size) and can be confusing; please differentiate these symbols clearly.
  7. [Sec. 4.2.1] The curriculum description says Phase 0 uses 'a single randomly sampled hair latent' but later phases use 1,000 and 20,000 latents; please clarify how many unique latents are used in each phase and how they are sampled.

Circularity Check

2 steps flagged · score 4.0 of 10

Eq. (4) is vacuous by construction (MSE of a function with itself) and Table 1's headline metrics re-measure the same physics losses used for training, so the physics-based fidelity claim is partly self-defined; the core drape prediction is nevertheless a genuine self-supervised generalization, not a fitted echo of a target output.

  1. self definitional [Section 3.3.4, Eq. (4)]
    "Specifically, we enforce that its functional behavior on the original static features H(t) static remains similar to that of the (frozen) static head: L(t) func_reg = λfr · MSE( F dyn pred(H(t) static), F dyn pred(H(t) static) ) (4)"

    Both arguments of the MSE are the same tensor, so L_func_reg is identically zero for every possible network weight and can never influence training. The prose says the loss maintains consistency between the static and dynamic prediction heads, but the printed equation compares F_dyn_pred with itself, not with F_static_pred. The functional-regularization mechanism claimed in Contribution (4) and credited in the Sec. 4.5.2 ablation therefore reduces to a tautology: by construction the constraint imposes nothing, and any static-feature preservation attributed to it is not produced by this equation.

  2. self definitional [Section 4.3, Table 1 vs Sections 3.4.1 and 3.4.4]
    "The total loss for the static network evaluated at timestep t is: Lstatic(t) tot = L(t)s + L(t)b + L(t)aux + L(t)smooth + L(t)gravity + L(t)hb + L(t)hh + L(t)root (5) ... As shown in Table 1, HairFormer significantly outperforms baselines in both static and dynamic performance. Method Time (ms/drape)↓ Penetration (%)↓ Length Change (%)↓ ... HairFormer (Dynamic) 12.197 6.751 1.752"

    The two headline metrics are direct numerical proxies for the training objectives: Penetration (%) is the fraction of hair vertices violating the signed-distance barrier that L_hb penalizes in Eq. (12), and Length Change (%) is the percentage version of the segment-length deviation penalized by L_s in Eq. (7). The network is trained by minimizing exactly these energies, and the Adam/L-BFGS baselines in Sec. 4.3.2 also minimize the same loss. Thus Table 1 largely certifies that the method minimized its own objective rather than matching an independent physics reference simulator or measured hair data. Physical plausibility is effectively defined as low values of these losses and then re-reported as evidence, making the quantitative validation partially circular by construction.

full rationale

The paper does not rely on a self-citation chain or an imported uniqueness theorem; the Transformer-based static and dynamic networks are trained self-supervised on AMASS pose sequences with hand-specified energy terms, and the comparison against an adapted Quaffure and direct optimization is a genuine engineering benchmark. However, two internal steps undermine the claimed derivation. First, Eq. (4) as printed is vacuously zero because it computes MSE between F_dyn_pred(H_static) and itself, so the functional-regularization mechanism intended to preserve static features does not exist as written. Second, the quantitative evaluation in Table 1 uses Penetration (%) and Length Change (%), which are near-proxies of the contact and inextensibility losses used for training, and the direct-optimization baselines minimize the same losses; consequently the reported numbers partly re-measure the optimized objective rather than validating physical fidelity against external ground truth. The limitations section further concedes residual penetrations around the breast area and hand joints, qualifying the abstract's claim of effectively resolving penetrations. These are fixable internal gaps rather than a wholesale equivalence of prediction and input, so the circularity is partial rather than total.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central method rests on manually weighted physics-inspired losses, the PERM/SMPL input domains, and a corrected interpretation of Eq. (4). No genuinely new physical entities are introduced; the main observational burden is the unvalidated equivalence between the chosen loss energies and real hair behavior.

free parameters (5)
  • Physics loss weights k_s, k_b, lambda_aux, lambda_smooth, lambda_hb, lambda_hh, lambda_root, lambda_inertia… = not reported
    All balancing coefficients in Eqs. (5), (6), and (7)-(16) are chosen by hand; the paper does not report their values, and the reported drapes and dynamics depend on them.
  • Vertex mass m_s,v' = not reported
    Mass per moving vertex enters the gravity and inertia losses (Eqs. 11 and 16); values are hardcoded and may be scaled for guide strands, but the exact values are not given.
  • Contact barrier parameters xi, d_hat, b_p, t0 = hair diameter or 1 mm; d_hat = 1.5x; b_p annealed from 0.9 to 0.01
    The extended C-IPC barrier in Eq. (12) depends on chosen hard and soft barriers and a deep-penetration anchor; these parameters shape the penetration-resolution behavior.
  • Dynamic-score filtering threshold = median of 500 sampled sequences
    Pose sequences with below-median dynamic score are discarded from dynamic training (Sec. 4.2.2), a data-selection choice that affects the learned dynamics.
  • Curriculum schedule and augmentation noise = not reported
    Static training uses a four-phase curriculum with body-shape Gaussian noise (Sec. 4.2.1); the noise scale and phase durations are not reported.
assumptions (4)
  • domain assumption The potential-energy losses in Sec. 3.4 are a faithful and sufficient model of real hair physics for drape and motion.
    The network is trained entirely by minimizing these losses; no reference simulation or measured hair data is used to validate that the minimized energy corresponds to true hair behavior.
  • domain assumption PERM StyleGAN hair latents and the fixed scalp UV root grid cover the space of 'arbitrary hairstyles' claimed in the paper.
    All inputs are generated from a pretrained PERM generator and roots are placed on a fixed 32x32 scalp grid (Sec. 3.1); bodies with different scalp geometry require recomputation, which the authors note.
  • domain assumption SMPL body parameters and AMASS pose sequences are sufficient to represent diverse body shapes and motions.
    Training and evaluation use SMPL and AMASS only (Secs. 3.1 and 4.1); no real scanned bodies or motions outside this parameterization are tested.
  • ad hoc to paper The functional-regularization loss as written in Eq. (4) is assumed to be a typo and implementable as a comparison between static and dynamic prediction heads.
    In the printed equation both arguments to the MSE are identical, so the term is zero by construction; the text describes a meaningful regularization, requiring an unstated correction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HairFormer: Transformer-Based Dynamic Neural Hair Simulation." pith.science (2026). https://pith.science/paper/4BBNVXBT

@misc{pith2026250712600,
  author       = {Pith},
  title        = {Pith review of: HairFormer: Transformer-Based Dynamic Neural Hair Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4BBNVXBT}},
  note         = {Machine review of arXiv:2507.12600}
}
read the original abstract

Simulating hair dynamics that generalize across arbitrary hairstyles, body shapes, and motions is a critical challenge. Our novel two-stage neural solution is the first to leverage Transformer-based architectures for such a broad generalization. We propose a Transformer-powered static network that predicts static draped shapes for any hairstyle, effectively resolving hair-body penetrations and preserving hair fidelity. Subsequently, a dynamic network with a novel cross-attention mechanism fuses static hair features with kinematic input to generate expressive dynamics and complex secondary motions. This dynamic network also allows for efficient fine-tuning of challenging motion sequences, such as abrupt head movements. Our method offers real-time inference for both static single-frame drapes and dynamic drapes over pose sequences. Our method demonstrates high-fidelity and generalizable dynamic hair across various styles, guided by physics-informed losses, and can resolve penetrations even for complex, unseen long hairstyles, highlighting its broad generalization.

Figures

Figures reproduced from arXiv: 2507.12600 by the authors.

Figure 1
Figure 1. The HairFormer pipeline for dynamic neural hair simulation: (a) Given an arbitrarily generated hairstyle, body shape, and a body pose (or pose sequence), (b) the hair is first rigidly positioned on the body. (c) A Transformer-powered static network then predicts the static draped shape for single frames, effectively resolving hair-body penetrations. (d) Finally, a dynamic network leverages the static predictions to … view at source ↗
Figure 2
Figure 2. Method Overview. Our network architecture features distinct static and dynamic modules. (Top right) Initially, hair strands are generated from hairstyle latent code using a StyleGAN-based model (PERM [He et al. 2024]) and then rigidly posed onto the character’s body parameterized by SMPL [Loper et al. 2023]. (Top left) The static network module processes the hairstyle latent and the body parameters to predict an ini… view at source ↗
Figure 3
Figure 3. Extrapolated barrier function of various orders. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Animated sequence of whiplash frames. (Requires a JavaScript [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Rigidly-posed hair versus simulated hair in Phase 0 training of the [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 8
Figure 8. Figure 8: Left to right: Rigidly-posed unseen hairstyles, Quaffure inference, [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Per-frame inference using our static network for an unseen hairstyle. Our network correctly predicts the back hair splitting, while maintaining [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Top: Static per-frame inference. Bottom: Dynamic inference for entire sequence. Red arrows indicate regions with significant differences. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 11 canonical work pages

  1. [6]

    Artur Grigorev, Michael J Black, and Otmar Hilliges

    Functional regularization for representation learning: A unified theoretical perspective.Advances in Neural Information Processing Systems 33 (2020), 17187–17199. Artur Grigorev, Michael J Black, and Otmar Hilliges

    work page 2020

  2. [8]

    arXiv preprint arXiv:2412.17144 (2024)

    Augmented Mass-Spring model for Real-Time Dense Hair Simulation. arXiv preprint arXiv:2412.17144 (2024). Geoffrey Hinton, Oriol Vinyals, and Jeff Dean

    arXiv 2024

  3. [10]

    ACM Transactions on Graphics (TOG) 43, 4 (2024), 1–11

    Real-time Physically Guided Hair Interpolation. ACM Transactions on Graphics (TOG) 43, 4 (2024), 1–11. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al

    work page 2024

  4. [11]

    ICLR 1, 2 (2022),

    Lora: Low-rank adaptation of large language models. ICLR 1, 2 (2022),

    work page 2022

  5. [12]

    arXiv preprint arXiv:2012.04457 (2020)

    Codimensional incremental potential contact. arXiv preprint arXiv:2012.04457 (2020). Hai Chuan Liu, Anis Salwa Mohd Khairuddin, Joon Huang Chuah, Xian Min Zhao, Xiao Dan Wang, and Li Ming Fang

    arXiv 2020

  6. [13]

    IEEE Access (2024)

    HCMT: A Novel Hierarchical Cross-modal Transformer for Recognition of Abnormal Behavior. IEEE Access (2024). Pengyu Long, Zijun Zhao, Min Ouyang, Qingcheng Zhao, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu

    work page 2024

  7. [15]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Snug: Self-supervised neural dynamic garments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8140–8150. Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lem- pitsky, and Egor Zakharov. 2023a. Neural haircut: Prior-guided strand-based hair reconstruction. In Proceedings of the IEEE/CVF Intern...

    arXiv 2023

  8. [16]

    Quaffure: Real-Time Quasi-Static Neural Hair Simulation

    Quaffure: Real-Time Quasi-Static Neural Hair Simulation. arXiv preprint arXiv:2412.10061 (2024). A Vaswani

    work page Pith review arXiv 2024

Show all 19 references
  1. [18]

    ACM Transactions on Graphics (TOG) 42, 6 (2023), 1–16

    Groomgen: A high-quality generative hair model using hierarchical latent represen- tations. ACM Transactions on Graphics (TOG) 42, 6 (2023), 1–16. , Vol. 1, No. 1, Article . Publication date: September

    2023

  2. [19]

    Bottom: Dynamic inference for entire sequence

    Top: Static per-frame inference. Bottom: Dynamic inference for entire sequence. Red arrows indicate regions with significant differences. , Vol. 1, No. 1, Article . Publication date: September 2025

    2025

  3. [2008]

    In ACM SIGGRAPH 2008 papers

    Discrete elastic rods. In ACM SIGGRAPH 2008 papers. 1–12. Hugo Bertiche, Meysam Madadi, and Sergio Escalera

    2008

  4. [2014]

    arXiv preprint arXiv:1412.3555 (2014)

    Empir- ical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014). Gilles Daviet

    2014 arXiv

  5. [2015]

    arXiv preprint arXiv:1503.02531 (2015)

    Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015). Jerry Hsu, Tongtong Wang, Zherong Pan, Xifeng Gao, Cem Yuksel, and Kui Wu

    2015 arXiv

  6. [2017]

    Advances in Neural Information Processing Systems (2017)

    Attention is all you need. Advances in Neural Information Processing Systems (2017). Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner

    2017

  7. [2020]

    arXiv preprint arXiv:2012.11310 (2020)

    Pbns: Physically based neural simulator for unsupervised garment pose space deformation. arXiv preprint arXiv:2012.11310 (2020). Hugo Bertiche, Meysam Madadi, and Sergio Escalera

    2020 arXiv

  8. [2022]

    ACM Transactions on Graphics (TOG) 41, 6 (2022), 1–14

    Neural cloth simulation. ACM Transactions on Graphics (TOG) 41, 6 (2022), 1–14. Paul J Besl and Neil D McKay

    2022

  9. [2023]

    In ACM SIGGRAPH 2023 Conference Proceedings

    Interactive Hair Simulation on the GPU using ADMM. In ACM SIGGRAPH 2023 Conference Proceedings . 1–11. Siddhant Garg and Yingyu Liang

    2023

  10. [2024]

    arXiv preprint arXiv:2407.19451 (2024)

    Perm: A parametric representation for multi-style 3d hair modeling. arXiv preprint arXiv:2407.19451 (2024). Jorge Alejandro Amador Herrera, Yi Zhou, Xin Sun, Zhixin Shu, Chengan He, Sören Pirk, and Dominik L Michels

    2024 arXiv

  11. [2025]

    arXiv preprint arXiv:2502.06392 (2025)

    TANGLED: Generating 3D Hair Strands from Images with Arbitrary Styles and Viewpoints. arXiv preprint arXiv:2502.06392 (2025). Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.