Pith. sign in

REVIEW 4 major objections 6 minor 75 references

From multi-view video, DynHair builds head avatars whose hair strands deform from head velocity, acceleration, and gravity rather than a hand-tuned simulator.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 14:02 UTC pith:ILGRRFGA

load-bearing objection Solid person-specific systems paper on strand-Gaussian hair driven by head kinematics; useful, not airtight on the dynamics claim. the 4 major comments →

arxiv 2607.23861 v1 pith:ILGRRFGA submitted 2026-07-26 cs.CV cs.GR

Head Avatars with Dynamic Explicit Hair

Vanessa Sklyarova , Haonan Chen , Berna Kabadayi , Tobias Kirschstein , Zicong Fan , Xi Wang , Gerard Pons-Moll , Matthias Nie{\ss}ner
show 3 more authors
Marc Pollefeys Michael J. Black Justus Thies
This is my paper
classification cs.CV cs.GR
keywords dynamic hair modelinghead avatarsstrand-based hair3D Gaussian Splattinghair trackingtemporal deformationdigital humans
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most photoreal head avatars treat hair as part of a blurry surface and ignore inertia and gravity, so the hair looks glued to the skull when the head turns or nods. DynHair instead rebuilds hair as explicit strands carrying 3D Gaussians, then learns a small temporal network that, given only a short history of the head’s angular velocity, acceleration, and gravity in head-local coordinates, predicts physically plausible per-point displacements from a fixed canonical hairstyle. The hair and face are trained together under image, silhouette, orientation, penetration, and elastic losses, yielding both tracked hair on the input video and an animatable avatar whose hair can be driven by novel head motion. The result is sharper strand detail, less temporal flicker, and better perceptual hair realism than unstructured Gaussian avatars and a hand-tuned physics simulator, while still allowing simple edits such as trim and brightness change.

Core claim

The paper shows that dynamic, person-specific hair for a Gaussian head avatar can be learned end-to-end from multi-view video by representing hair as strand-aligned Gaussians whose non-rigid motion is predicted relative to a canonical style from head-local angular velocity, acceleration, and relative gravity alone, jointly optimized with the face under photometric, geometric, and elastic supervision.

What carries the argument

The hair deformation network: an LSTM encodes a sliding window of head-local angular velocity, acceleration, and relative gravity into a motion embedding; FiLM then modulates per-point strand features, and an MLP with learnable root-to-tip attenuation predicts displacements of the canonical strands.

Load-bearing premise

A short window of the head’s own spin rate, spin-up, and gravity direction is enough to determine how that person’s hair should move, without needing the previous hair shape or explicit collisions with the body.

What would settle it

Drive the trained model with held-out head motions that include strong hair–shoulder contact or long free-swing sequences, then check whether hair velocity, silhouette IoU, and temporal LPIPS still match the multi-view ground truth better than a simulator and unstructured Gaussian baselines; clear failure under body contact or accumulating drift would refute the claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Animatable head avatars can carry controllable, strand-level hair dynamics driven only by tracked head pose history.
  • Hair appearance and geometry stay editable after training (trim, recolor) while the learned dynamics still apply.
  • Perceptual and temporal hair metrics improve even when pixel-aligned scores stay competitive with unstructured Gaussians.
  • The same motion conditioning transfers across subjects in cross-reenactment without retuning physics parameters.
  • Canonical-relative prediction avoids the long-horizon drift of frame-to-frame hair-state predictors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the head-motion-only prior is truly sufficient for many everyday hairstyles, large multi-view datasets could train a single generalist dynamics head that personalizes with few shots rather than per-subject training from scratch.
  • Explicit strand Gaussians plus elastic loss may transfer to other thin deformable structures (cloth fringes, fur) where unstructured Gaussians over-smooth.
  • Adding an explicit hair–body collision term on top of the learned prior is the natural next stress test once full-body multi-view data is available.
  • The gap between FID/temporal metrics and PSNR/SSIM suggests evaluation suites for dynamic hair should weight motion plausibility over single-frame pixel match.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper presents DynHair, a person-specific head-avatar method that disentangles hair from the face/upper body. Hair is represented as explicit polyline strands (initialized from the Im2Haircut prior) carrying strand-aligned 3D Gaussians; a temporal deformation network — an LSTM over a sliding window of head-local angular velocity, angular acceleration, and relative gravity, FiLM-modulating per-point strand features before an MLP — predicts per-frame non-rigid displacements from a canonical hairstyle. Training is end-to-end through differentiable splatting with photometric, silhouette/orientation, penetration, and elastic losses, jointly with a GHA-style face/body model. Experiments on three HHAvatar scenes and two newly captured 15-camera scenes compare against GaussianAvatars, Gaussian Head Avatar, and a hand-tuned Maya simulator (with learned colors transferred), using reconstruction metrics, a new temporal metric (tLPIPS_ex), hair-motion statistics (velocity/acceleration), physics-inspired metrics (VER, CTS, P95-Disp, AMD), an ablation study, and a 16-person user study. The claimed contributions are improved hair dynamics, temporal consistency, and cross-subject generalization, plus an animatable, editable strand-based avatar.

Significance. If the results hold, this is a meaningful step: it is, to my knowledge, the first method to learn strand-level hair dynamics end-to-end from multi-view video within a Gaussian head-avatar framework, replacing hand-tuned simulation parameters with a data-driven deformation model regularized by physically motivated losses (elastic segment preservation, root attenuation, gravity/velocity/acceleration conditioning). The explicit strand representation yields practical benefits (temporal correspondence by construction, trimming/recoloring that propagates over sequences, Sec. D). The authors commit to releasing code and data, including a new 15-camera, 72 FPS, 4K multi-view dataset with a scripted motion protocol, which is a genuine community asset. The ablations (Table 2) support the individual design choices, and the tLPIPS_ex metric is a reasonable, clearly-defined temporal measure. These strengths, however, are currently established against baselines that do not model hair dynamics at all (GHA, GA) or a hand-tuned simulator, which limits the strength of the "state-of-the-art" claim.

major comments (4)
  1. [Sec. 3.1, Eqs. (2)-(7); Sec. 4] Eqs. (2)-(7), Sec. 4 ('window history of 5 frames'): the deformation is a deterministic function of the canonical hairstyle plus a ~70 ms window (5 frames at 72 FPS) of head-local omega, alpha, g. Physically distinct hair states can share indistinguishable conditioning windows: after a head shake stops, omega and alpha decay to ~0 within a few frames while real hair continues swinging/settling for ~0.5-2 s. Two test frames with identical recent head kinematics -- one mid-settle, one at rest -- receive the same z_t and hence the same predicted displacement. The network can then only fit the conditional mean over such states, biasing it toward under-motion precisely in the follow-through/settling regime the Introduction highlights as the contribution. The paper itself (Sec. 2, Sec. E.4) notes HHAvatar/HADES condition on past hair states for this reason, and offers only a drift-based hypoth
  2. [Sec. 4 (Dataset details) and Sec. 4.1 evaluation protocol] The test split is a temporal split of the same scripted 22-action protocol (Sec. 4, 'All participants performed the same 22 actions'). With deterministic action categories repeated per subject, the model can succeed by recognizing action patterns from the conditioning window rather than inferring dynamics from physics. A concrete, inexpensive check: hold out entire action categories (e.g., train on 19, test on impulse/figure-eight) or at minimum report metrics separately for the 10 hair-dynamics actions vs. the rest. As it stands, the generalization claim ('generalization across subjects and motion patterns') is supported only for within-protocol interpolation, and cross-reenactment (Fig. 4) is qualitative only.
  3. [Sec. 4.2, Eq. (18) (AMD) and Table 1 (vel. column)] Eq. (18) defines AMD as deviation of total hair angular momentum from its frame-1 value, presented as 'violations of angular momentum conservation' with lower = 'more physically plausible'. But the hair is not an isolated system: it is externally forced by the head (which the measurement itself removes only partially by unposing) and subject to gravity and air drag, so its angular momentum should not be conserved; a simulator that conserves it would be wrong. The Maya reference row (AMD 7120) is then not a meaningful physical anchor. Relatedly, the Table 1 'vel.' metric has no ground-truth anchor (no GT 3D hair motion exists), so 'higher velocity = more expressive' conflates expressiveness with jitter; the justification given ('accel. measures temporal smoothness') does not resolve this. These two metrics carry part of the 'physical realism' claim of Table 2/Table 1; they need either a p
  4. [Sec. 4.1 (Baselines) and abstract SOTA claim] The stated motivation is that capture-based dynamic-hair methods are the relevant prior art, yet neither HHAvatar nor HADES is evaluated against -- HHAvatar is not run even though its dataset (3 of the 5 scenes) is used for training/evaluation. The physics baseline is a hand-tuned Maya simulation of 200 sparse strands (16 points, interpolated via kNN to the full hairstyle) with colors transferred from the authors' own model; this is a weak appearance baseline and its physics parameters (Sec. C.1) are admittedly not fitted. The remaining baselines (GHA, GA) do not model hair dynamics at all. Given this, 'state-of-the-art performance in terms of hair dynamics' (abstract) is stronger than the evidence supports. Either add a comparison to HHAvatar on its own data (even with a third-party reimplementation) or temper the claim to 'outperforms general head-avatar baselines and a hand-tuned simu
minor comments (6)
  1. [Sec. 3.1, Eq. (2)] Eq. (2) uses central differences, requiring R_{t+1}; the conditioning is therefore non-causal (the 'history' contains future information). This partially mitigates the short-window concern and should be stated explicitly, including how boundary frames and real-time inference are handled.
  2. [Eq. (12) (orientation loss) and tLPIPS_ex definition] The wrap-aware min over {theta, theta-1, theta+1} implies angles normalized to period 1, but the loss is multiplied by pi; please state the angle convention explicitly. Also define the clamp operator in the tLPIPS_ex mask equation and the sign convention (positive = excess variation) inline at first use rather than only in the surrounding text.
  3. [Table 2] All ablations are averaged over only 2 scenes and 3 motion types at 240k iterations (vs. 320k for the main model); please confirm conclusions hold at full training and report per-scene variance, since Table 2 differences (e.g., tLPIPS_ex 0.0024 vs. 0.0028 for 'w/ pose') are within plausible seed/scene noise.
  4. [Sec. A.2 / B.1] Hair Gaussian opacity is fixed to 1 (Sec. A.2/B.1); please comment on whether this limits wispy/semi-transparent regions, and report the ~21 FPS inference number (Sec. B.4) in the main paper alongside the 5-day training cost.
  5. [Table 3 and Sec. B.1] The user study (16 participants) omits GA without comment; the static loss uses lambda_pen while Eq. (10) uses lambda_penetr -- unify notation; Sec. B.1 states strands 'have 200 points' then are resampled to 40, clarify this is intentional downsampling.
  6. [Figs. 3-5] Figures 3/4/9/10 would benefit from temporal strips or heatmaps of per-frame hair-region error; single frames under-support the temporal-coherence claims, which are the paper's core contribution.

Circularity Check

0 steps flagged

No significant circularity: DynHair is an empirical learned avatar model whose dynamics are fit to video and scored on held-out motions, not a first-principles derivation that collapses to its inputs.

full rationale

The paper's load-bearing chain is standard supervised/self-supervised learning, not a claimed algebraic or uniqueness derivation. Canonical strands H0 are initialized from a prior (Im2Haircut) and then jointly refined; non-rigid displacements are the free outputs of an LSTM+FiLM+MLP conditioned on measured head kinematics (Eqs. 2–7), optimized under photometric, silhouette/orientation, penetration, elastic, and color losses on multi-view training frames; evaluation is on held-out temporal splits, cross-reenactment, and external baselines (GHA, GaussianAvatars, hand-tuned Maya). Nothing in that chain defines the predicted quantity in terms of itself, renames a fitted constant as a prediction, or rests the central dynamics claim on an unverified self-citation uniqueness theorem. Self-citations (Im2Haircut, Neural Haircut, PhysHead, GHA-style body) supply initialization, related work, or the face backbone; they do not force the reported hair-dynamics result. The skeptic's 5-frame conditioning/identifiability concern is a modeling-capacity limitation, not circularity under the rubric. Score 0 with empty steps is the honest finding.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

Central claim rests on standard differentiable rendering and face tracking plus several modeling choices treated as given: strand polylines as the right hair state, head rigid-body signals as sufficient motion drivers, Im2Haircut canonical geometry, GHA-style face Gaussians, and a stack of hand-set loss weights and architectural hyperparameters. No new physical law is postulated; the 'physics' is soft regularization and input features, not a calibrated simulator.

free parameters (6)
  • Loss weights (λ_rgb, λ_ssim, λ_vgg, λ_seg, λ_orient, λ_penetr, λ_elastic, color regs) = e.g. λ_elastic=500, λ_seg=10 with decay, λ_vgg 0.1→0.6 (App. B.1)
    Hand-set schedule including λ_elastic=500 and late λ_vgg boost; directly shapes the claimed dynamics vs appearance tradeoff.
  • Motion history window T = T=5
    Sliding window length for LSTM conditioning; fixed to 5 frames in all experiments.
  • Strand discretization (N strands, L points) = L=40, ~11k strands
    Resolution of the explicit hair state; affects both metrics and capacity.
  • Learnable root attenuation ρ_j = linear ramp init to 1.0 at tip
    Initialized as linear ramp root→tip; encodes less root motion by construction and is optimized.
  • Per-axis ω/α normalization maxima from training set = dataset-dependent max |ω|, |α|
    Clamps motion inputs to [-1,1] using training statistics; couples conditioning scale to the captured motions.
  • Maya baseline physical parameters = stretchResistance=600, mass=2, drag=0.65, etc. (App. C.1)
    Hand-tuned simulator settings used as a dynamics baseline; weakens the simulator comparison if mistuned.
axioms (6)
  • domain assumption Differentiable 3D Gaussian splatting is a faithful enough image formation model for joint photometric and silhouette supervision of hair and face.
    Used throughout §3 for end-to-end training; standard in the 3DGS avatar literature.
  • domain assumption BFM/multi-view 3DMM tracking supplies accurate enough global head pose to derive angular velocity, acceleration, and head-local gravity.
    §3.1 motion conditioning and App. B.2; limitations admit tracker failure under occlusion and extreme yaw.
  • ad hoc to paper A short history of head-local ω, α, and g is a sufficient statistic for person-specific non-rigid hair displacement from a canonical style.
    Core modeling choice in §3.1; contrasts with methods that condition on past hair state (HHAvatar, HADES).
  • domain assumption Im2Haircut PCA prior plus multi-view static optimization yields a usable canonical strand geometry including occluded inner structure.
    §3.1 Canonical representation; errors here propagate to all dynamic frames.
  • ad hoc to paper Elastic segment-length preservation and root attenuation are adequate physics proxies without full collision, friction, or material identification.
    L_elastic and ρ_j in §3.1–3.2; limitations note hair–body collisions are not explicit.
  • domain assumption Pretrained hair/body segmentation and Gabor orientation maps are accurate enough geometric supervisors.
    L_seg, L_orient; limitations cite mask errors and clothing/hair bleed.
invented entities (2)
  • DynHair hair deformation network (LSTM on γ(ω,α,g) → FiLM → strand MLP with root attenuation) no independent evidence
    purpose: Map head motion history to per-point displacements of canonical strands.
    Architectural construct trained per subject; not a new physical entity, but the paper's main postulated mechanism for dynamics.
  • Custom temporal/physics metrics (tLPIPS_ex, VER, CTS, P95-Disp, AMD) no independent evidence
    purpose: Quantify temporal flickering and physical plausibility beyond PSNR/SSIM.
    Defined in §4; useful but paper-specific, so they partly measure what the representation is built to produce.

pith-pipeline@v1.2.0-grok45-kimik3 · 27487 in / 4223 out tokens · 90028 ms · 2026-07-30T14:02:22.624745+00:00 · methodology

0 comments
read the original abstract

We present DynHair, a novel method for tracking and modeling dynamic hair for human head avatars. From video input, we reconstruct a dynamic head avatar with an explicit strand-based hair representation using structured 3D Gaussian Splatting. In contrast to the face region of human head avatars, which can be modeled with 3D Gaussians that are attached or generated with respect to some expressive 3D head model, hair is particularly challenging as it exhibits dynamic motion effects. Therefore, we present a novel method that models the dynamic deformations of the hair strands using a temporal network that is conditioned on angular velocity and acceleration of the head, as well as relative gravity. Specifically, an LSTM encodes the motion history and modulates per-point strand features via FiLM conditioning which further used by MLP to produce physically plausible displacements to canonical hairstyle. We jointly optimize this motion and appearance representation of the hair, with a 3DGS-based representation of the face-region, via differentiable Gaussian splatting with photometric, geometric, and physics-based supervision. As a result of our method, we retrieve hair tracking of the training video data and an animatable head avatar with controllable hair dynamics. In our experiments, we demonstrate state-of-the-art performance in terms of hair dynamics, temporal consistency, and generalization across subjects.

Figures

Figures reproduced from arXiv: 2607.23861 by Berna Kabadayi, Gerard Pons-Moll, Haonan Chen, Justus Thies, Marc Pollefeys, Matthias Nie{\ss}ner, Michael J. Black, Tobias Kirschstein, Vanessa Sklyarova, Xi Wang, Zicong Fan.

Figure 1
Figure 1. Figure 1: DynHair is a novel method for learning hair dynamics of 3D strands from data. It leverages multi-view video input to learn a dynamic hair strand deformation model based on the velocity and acceleration of the human head. Abstract. We present DynHair, a novel method for tracking and mod￾eling dynamic hair for human head avatars. From video input, we recon￾struct a dynamic head avatar with an explicit strand… view at source ↗
Figure 2
Figure 2. Figure 2: Method overview. Given multi-view video input, we first initialize a canon￾ical hybrid model for the upper body and hair. Hair strands are initialized using the Im2Haircut [52] prior model, and the upper body is represented with unstructured 3D Gaussians [64]. We then jointly optimize hair and head deformations with two separate motion models. The hair deformation is conditioned on BFM motion history with … view at source ↗
Figure 3
Figure 3. Figure 3: Self-reenactment comparison. Each row shows a different test frame or subject. Columns from left to right: ground-truth image, physics-based simulator from Maya [1], GaussianAvatars [47], Gaussian Head Avatar [64], and our method. Our ap￾proach produces sharper strand-level detail and more temporally coherent hair motion. proach against state-of-the-art multi-view dynamic head avatar methods, Gaus￾sian Hea… view at source ↗
Figure 4
Figure 4. Figure 4: Cross-reenactment comparison. The 1st row shows left-right head rotation, while the second row corresponds to nodding. Our hairstyle better follows physics constraints and gravity. The image in the first column serves as a reference for human motion. While the hairstyle should follow general dynamics, it is not an exact match. neighbor displacement between consecutive hair point clouds in this head-local f… view at source ↗
Figure 5
Figure 5. Figure 5: We show more results on newly captured sequences with diverse hair motions. Notably, ours recovers per-strand temporal hair motion. We include an ablation without the non-rigid deformation module, where hair moves rigidly with the head transform. the position of the j-th point on the i-th strand at frame t, and ∆t the time step. Then velocity explosion ratio (VER) is defined as the ratio of maximum to medi… view at source ↗
Figure 6
Figure 6. Figure 6: Hairstyle manipulation. The reconstructed dynamic hairstyle can be further manipulated by changing the brightness of Gaussians or trimming them [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation on color regularization and increased weight for the VGG loss. E.3 Limitations Despite producing plausible dynamic hair geometry and appearance, our method has several limitations. Hair-face boundary artifacts. Since hair and face are reconstructed with separate representations, strand-based Gaussians for hair and mesh-anchored Gaussians for the face, visible artifacts can occur at their boundary.… view at source ↗
Figure 8
Figure 8. Figure 8: Limitations of our method: tracking errors of the face caused by occlusions (left) and hair structure in regions where the segmentation masks have artifacts [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Extended comparison with baselines on new captures [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Extended comparison with baselines on new captures [PITH_FULL_IMAGE:figures/full_fig_p032_10.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

75 extracted references · 1 linked inside Pith

  1. [1]

    Autodesk, INC. maya. Maya, accessed: 2025-10-10

    2025

  2. [2]

    Unreal Engine, accessed: 2025-10-10

    Epic Games unreal engine. Unreal Engine, accessed: 2025-10-10

    2025

  3. [3]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Aneja, S., Sevastopolsky, A., Kirschstein, T., Thies, J., Dai, A., Nießner, M.: Gaus- sianspeech: Audio-driven personalized 3d gaussian avatars. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13065– 13075 (October 2025)

    2025

  4. [4]

    In: Musialski, P., Lim, I

    Ben Ayed, F., Becherini, G., Thies, J., Sklyarova, V.: Vid2Haircut: 3D Strand- Based Hairstyle Reconstruction from Video. In: Musialski, P., Lim, I. (eds.) Euro- graphics 2026 - Short Papers. The Eurographics Association (2026)

    2026

  5. [5]

    Transactions on Graphics, (Proc

    Bertails, F., Audoly, B., Cani, M.P., Querleux, B., Leroy, F., Lévêque, J.L.: Super- helices for predicting the dynamics of natural hair. Transactions on Graphics, (Proc. SIGGRAPH)25(3), 1180–1187 (2006)

    2006

  6. [6]

    In: Computer graphics forum

    Blanz, V., Basso, C., Poggio, T., Vetter, T.: Reanimating faces in images and video. In: Computer graphics forum. vol. 22, pp. 641–650. Wiley Online Library (2003)

    2003

  7. [7]

    Bouaziz, S., Wang, Y., Pauly, M.: Online modeling for realtime facial animation 32(4) (Jul 2013)

    2013

  8. [8]

    In: International Conference on Computer Vision (ICCV) (2021)

    Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In: International Conference on Computer Vision (ICCV) (2021)

    2021

  9. [9]

    In: SIGGRAPH Conference Papers (SA)

    Chen, Y., Wang, L., Li, Q., Xiao, H., Zhang, S., Yao, H., Liu, Y.: Monogaussiana- vatar: Monocular gaussian point-based head avatar. In: SIGGRAPH Conference Papers (SA). pp. 1–9 (2024)

    2024

  10. [10]

    In: The Thirty-eighthAnnualConferenceonNeuralInformationProcessingSystems(2024)

    Chu, X., Harada, T.: Generalizable and animatable gaussian head avatar. In: The Thirty-eighthAnnualConferenceonNeuralInformationProcessingSystems(2024)

    2024

  11. [11]

    Community, B.O.: Blender - a 3d modelling and rendering package (2018)

    2018

  12. [12]

    Daněček, R., Schmitt, C., Polikovsky, S., Black, M.J.: Supervising 3D talking head avatarswithanalysis-by-audio-synthesis.In:InternationalConferenceon3DVision (3DV) (2026)

    2026

  13. [13]

    In: SIGGRAPH Conference Papers (SA)

    Daviet, G.: Interactive hair simulation on the gpu using admm. In: SIGGRAPH Conference Papers (SA). SIGGRAPH ’23, Association for Computing Machinery, New York, NY, USA (2023)

    2023

  14. [14]

    arXiv preprint arXiv:2309.06441 (2023)

    Feng, Y., Liu, W., Bolkart, T., Yang, J., Pollefeys, M., Black, M.J.: Learning dis- entangled avatars with hybrid 3d representations. arXiv preprint arXiv:2309.06441 (2023)

    Pith/arXiv arXiv 2023

  15. [15]

    In: Conference on Computer Vision and Pattern Recognition (CVPR)

    Gafni, G., Thies, J., Zollhofer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649–8658 (2021)

    2021

  16. [16]

    ACM Transactions on Graphics (TOG)41(6), 1–12 (2022)

    Gao, X., Zhong, C., Xiang, J., Hong, Y., Guo, Y., Zhang, J.: Reconstructing per- sonalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG)41(6), 1–12 (2022)

    2022

  17. [17]

    In: Computer graphics forum

    Garbin, S.J., Kowalski, M., Estellers, V., Szymanowicz, S., Rezaeifar, S., Shen, J., Johnson, M.A., Valentin, J.: Voltemorph: Real-time, controllable and generalizable animation of volumetric representations. In: Computer graphics forum. vol. 43, p. e15117. Wiley Online Library (2024)

    2024

  18. [18]

    In: SIGGRAPH Asia 2024 Conference Papers

    Giebenhain, S., Kirschstein, T., Rünz, M., Agapito, L., Nießner, M.: Npga: Neural parametric gaussian avatars. In: SIGGRAPH Asia 2024 Conference Papers. pp. 1–11 (2024) Head Avatars with Dynamic Explicit Hair 17

    2024

  19. [19]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neu- ral head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18653–18664 (2022)

    2022

  20. [20]

    In: SIGGRAPH Asia 2024 Real-Time Live! SA Real-Time Live! ’24 (2024)

    He, C., Amador Herrera, J.A., Zhou, Y., Shu, Z., Sun, X., Feng, Y., Pirk, S., Michels, D.L., Zhang, M., Wang, T.Y., Rushmeier, H.: Digital salon: An ai and physics-driven tool for 3d hair grooming and simulation. In: SIGGRAPH Asia 2024 Real-Time Live! SA Real-Time Live! ’24 (2024)

    2024

  21. [21]

    In: International Conference on Learning Represen- tations (ICLR) (2025)

    He, C., Sun, X., Shu, Z., Luan, F., Pirk, S., Herrera, J.A.A., Michels, D.L., Wang, T.Y., Zhang, M., Rushmeier, H., Zhou, Y.: Perm: A parametric representation for multi-style 3d hair modeling. In: International Conference on Learning Represen- tations (ICLR) (2025)

    2025

  22. [22]

    In: International Conference on Computer Vision (ICCV) (2024)

    Herrera, J.A.A., Zhou, Y., Sun, X., Shu, Z., He, C., Pirk, S., Michels, D.L.: Aug- mented mass-spring model for real-time dense hair simulation. In: International Conference on Computer Vision (ICCV) (2024)

    2024

  23. [23]

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium (2018)

    2018

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Hong, Y., Peng, B., Xiao, H., Liu, L., Zhang, J.: Headnerf: A real-time nerf-based parametric head model. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 20374–20384 (2022)

    2022

  25. [25]

    Computer Graphics and Interactive Techniques (PACMCGIT)6(3) (2023)

    Huang, L., Yang, F., Wei, C., Chen, Y.J.E., Yuan, C., Gao, M.: Towards realtime: A hybrid physics-based method for hair animation on gpu. Computer Graphics and Interactive Techniques (PACMCGIT)6(3) (2023)

    2023

  26. [26]

    In: SIGGRAPH

    Iandola, F., Pidhorskyi, S., Santesteban, I., Gupta, D., Pahuja, A., Bartolovic, N., Yu, F., Garbin, E., Simon, T., Saito, S.: SqueezeMe: Mobile-ready distillation of gaussian full-body avatars. In: SIGGRAPH. ACM (2025)

    2025

  27. [27]

    ACM Trans

    Ichim, A.E., Bouaziz, S., Pauly, M.: Dynamic 3d avatar creation from hand-held video input. ACM Trans. Graph.34(4) (Jul 2015)

    2015

  28. [28]

    In: Proc

    Kabadayi, B., Sklyarova, V., Zielonka, W., Thies, J., Pons-Moll, G.: PhysHead: Simulation-ready gaussian head avatars. In: Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (June 2026)

    2026

  29. [29]

    In: International Confer- ence on 3D Vision (3DV) (March 2024)

    Kabadayi, B., Zielonka, W., Bhatnagar, B.L., Pons-Moll, G., Thies, J.: Gan-avatar: Controllable personalized gan-based human head avatar. In: International Confer- ence on 3D Vision (3DV) (March 2024)

    2024

  30. [30]

    In: AAAI (2022)

    Ke, Z., Sun, J., Li, K., Yan, Q., Lau, R.W.: Modnet: Real-time trimap-free portrait matting via objective decomposition. In: AAAI (2022)

    2022

  31. [31]

    Transactions on Graphics, (Proc

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. Transactions on Graphics, (Proc. SIGGRAPH) 42(4), 139–1 (2023)

    2023

  32. [32]

    ACM Transactions on Graphics (TOG)42(4), 1–14 (2023)

    Kirschstein, T., Qian, S., Giebenhain, S., Walter, T., Nießner, M.: Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG)42(4), 1–14 (2023)

    2023

  33. [33]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kirschstein, T., Romero, J., Sevastopolsky, A., Nießner, M., Saito, S.: Avat3r: Large animatable gaussian reconstruction model for high-fidelity 3d head avatars. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12089–12100 (2025)

    2025

  34. [34]

    Transactions on Graphics, (Proc

    Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4d scans. Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194–1 (2017)

    2017

  35. [35]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2025) 18 V

    Liao, Z., Tu, H., Peng, C., Zhang, H., Zhou, B., Liu, Y.: Hades: Human avatar with dynamic explicit hair strands. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2025) 18 V. Sklyarova et al

    2025

  36. [36]

    Transactions on Pattern Analysis and Machine Intelligence (TPAMI)PP(2025)

    Liao, Z., Xu, Y., Li, Z., Li, Q., Zhou, B., Bai, R., Xu, D., Zhang, H., Liu, Y.: Hha- vatar: Gaussian head avatar with dynamic hairs. Transactions on Pattern Analysis and Machine Intelligence (TPAMI)PP(2025)

    2025

  37. [37]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Liu, K., Choi, O., Wang, J., Hwang, W.: Cdgnet: Class distribution guided network for human parsing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4473–4482 (June 2022)

    2022

  38. [38]

    Transactions on Graphics, (Proc

    Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. Transactions on Graphics, (Proc. SIGGRAPH)40(4) (2021)

    2021

  39. [39]

    arXiv (2024)

    Luo, H., Ouyang, M., Zhao, Z., Jiang, S., Zhang, L., Zhang, Q., Yang, W., Xu, L., Yu, J.: GaussianHair: Hair modeling and rendering with light-aware gaussians. arXiv (2024)

    2024

  40. [40]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460–4470 (2019)

    2019

  41. [41]

    In: Eu- ropean Conference on Computer Vision (ECCV)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Eu- ropean Conference on Computer Vision (ECCV). Springer (2020)

    2020

  42. [42]

    In: DAGM German Conference on Pattern Recognition

    Nehvi, J., Kabadayi, B., Valentin, J., Thies, J.: Volumetric portrait avatar. In: DAGM German Conference on Pattern Recognition. pp. 3–19. Springer (2024)

    2024

  43. [43]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin- Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5865–5874 (2021)

    2021

  44. [44]

    ACM Trans

    Park,K.,Sinha,U.,Hedman,P.,Barron,J.T.,Bouaziz,S.,Goldman,D.B.,Martin- Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topo- logically varying neural radiance fields. ACM Trans. Graph.40(6) (dec 2021)

    2021

  45. [45]

    In: 2009 sixth IEEE interna- tional conference on advanced video and signal based surveillance

    Paysan, P., Knothe, R., Amberg, B., Romdhani, S., Vetter, T.: A 3d face model for pose and illumination invariant face recognition. In: 2009 sixth IEEE interna- tional conference on advanced video and signal based surveillance. pp. 296–301. Ieee (2009)

    2009

  46. [46]

    In: AAAI (2018)

    Perez, E., Strub, F., de Vries, H., Dumoulin, V., Courville, A.C.: Film: Visual reasoning with a general conditioning layer. In: AAAI (2018)

    2018

  47. [47]

    In: Con- ference on Computer Vision and Pattern Recognition (CVPR)

    Qian, S., Kirschstein, T., Schoneveld, L., Davoli, D., Giebenhain, S., Nießner, M.: Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. In: Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 20299–20309 (2024)

    2024

  48. [48]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    Rosu, R.A., Wu, K., Feng, Y., Zheng, Y., Black, M.J.: DiffLocks: Generating 3d hair from a single image using diffusion models. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    2025

  49. [49]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 130–141 (2024)

    2024

  50. [50]

    In: Conference on Computer Vision and Pattern Recognition (CVPR)

    Shao, Z., Wang, Z., Li, Z., Wang, D., Lin, X., Zhang, Y., Fan, M., Wang, Z.: Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1606–1616 (2024)

    2024

  51. [51]

    19705–19716 (2023) Head Avatars with Dynamic Explicit Hair 19

    Sklyarova, V., Chelishev, J., Dogaru, A., Medvedev, I., Lempitsky, V., Zakharov, E.:Neuralhaircut:Prior-guidedstrand-basedhairreconstruction.2023IEEE/CVF International Conference on Computer Vision (ICCV) pp. 19705–19716 (2023) Head Avatars with Dynamic Explicit Hair 19

    2023

  52. [52]

    In: International Conference on Computer Vision (ICCV) (2025)

    Sklyarova, V., Zakharov, E., Prinzler, M., Becherini, G., Black, M., Thies, J.: Im2haircut: Single-view strand-based hair reconstruction for human avatars. In: International Conference on Computer Vision (ICCV) (2025)

    2025

  53. [53]

    In: Conference on Com- puter Vision and Pattern Recognition (CVPR) (2024)

    Stuyck, T., Lin, G.W.C., Larionov, E., Chen, H.y., Bozic, A., Sarafianos, N., Roble, D.: Quaffure: Real-time quasi-static neural hair simulation. In: Conference on Com- puter Vision and Pattern Recognition (CVPR) (2024)

    2024

  54. [54]

    In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers

    Teotia, K., Rhodin, H., Mendiratta, M., Kim, H., Habermann, M., Theobalt, C.: Audio driven universal gaussian head avatars. In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers. SA Conference Papers ’25, Association for Comput- ing Machinery, New York, NY, USA (2025)

    2025

  55. [55]

    In: International Conference on 3D Vision 2025 (2025)

    Thambiraja, B., Prinzler, M., Aliakbarian, S., Cosker, D., Thies, J.: 3diFACE: Synthesizing and editing holistic 3d facial animation. In: International Conference on 3D Vision 2025 (2025)

    2025

  56. [56]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wang, C., Kang, D., Sun, H., Qian, S., Wang, Z., Bao, L., Zhang, S.H.: Mega: Hybrid mesh-gaussian head avatar for high-fidelity rendering and head editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 26274–26284 (June 2025)

    2025

  57. [57]

    IEEE Transactions on Visualization and Computer Graphics (2025)

    Wang, J., Xie, J.C., Li, X., Xu, F., Pun, C.M., Gao, H.: Gaussianhead: High- fidelity head avatars with learnable gaussian derivation. IEEE Transactions on Visualization and Computer Graphics (2025)

    2025

  58. [58]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2025)

    Wang,J.,Xu,Y.,Tretschk,E.,Wang,Z.,Ianina,A.,Bozic,A.,Neumann,U.,Tung, T.: Dgh: Dynamic gaussian hair. In: Advances in Neural Information Processing Systems (NeurIPS) (2025)

    2025

  59. [59]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)

    Wang, Z., Nam, G., Stuyck, T., Lombardi, S., Cao, C., Saragih, J., Zollhoefer, M., Hodgins, J., Lassner, C.: Neuwigs: A neural dynamic model for volumetric hair cap- ture and animation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)

    2023

  60. [60]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    Wang, Z., Nam, G., Stuyck, T., Lombardi, S., Zollhoefer, M., Hodgins, J., Lassner, C.: Hvh: Learning a hybrid neural volumetric representation for dynamic hair performance capture. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    2022

  61. [61]

    ACM transactions on graphics (TOG)30(4), 1–10 (2011)

    Weise, T., Bouaziz, S., Li, H., Pauly, M.: Realtime performance-based facial ani- mation. ACM transactions on graphics (TOG)30(4), 1–10 (2011)

    2011

  62. [62]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    Wu, K., Yang, L., Kuang, Z., Feng, Y., Han, X., Shen, Y., Fu, H., Zhou, K., Zheng, Y.: Monohair: High-fidelity hair modeling from a monocular video. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    2024

  63. [63]

    In: Conference on Computer Vision and Pattern Recognition (CVPR)

    Xiang, J., Gao, X., Guo, Y., Zhang, J.: Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1802–1812 (2024)

    2024

  64. [64]

    In: Conference on Computer Vision and Pattern Recognition (CVPR)

    Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1931–1941 (2024)

    1931

  65. [65]

    Transactions on Graphics, (Proc

    Yang, L., Shi, Z., Zheng, Y., Zhou, K.: Dynamic hair modeling from monocular videos using deep neural networks. Transactions on Graphics, (Proc. SIGGRAPH) 38(6) (2019)

    2019

  66. [66]

    In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)

    Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)

    2021

  67. [67]

    In: European Conference of Computer Vision (ECCV) (2024) 20 V

    Zakharov, E., Sklyarova, V., Black, M.J., Nam, G., Thies, J., Hilliges, O.: Human hair reconstruction with strand-aligned 3d gaussians. In: European Conference of Computer Vision (ECCV) (2024) 20 V. Sklyarova et al

    2024

  68. [68]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018)

    2018

  69. [69]

    In: Conference on Computer Vision and Pattern Recognition (CVPR)

    Zheng, Y., Abrevaya, V.F., Bühler, M.C., Chen, X., Black, M.J., Hilliges, O.: Im avatar: Implicit morphable head avatars from videos. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13545–13555 (2022)

    2022

  70. [70]

    Transactions on Graphics, (Proc

    Zhou, Y., Chai, M., Pepe, A., Gross, M., Beeler, T.: Groomgen: A high-quality generative hair model using hierarchical latent representations. Transactions on Graphics, (Proc. SIGGRAPH) (2023)

    2023

  71. [71]

    Transactions on Graph- ics, (Proc

    Zhou, Y., Chai, M., Wang, D., Winberg, S., Wood, E., Sarkar, K., Gross, M., Beeler, T.: Groomcap: High-fidelity prior-free hair capture. Transactions on Graph- ics, (Proc. SIGGRAPH) (2024)

    2024

  72. [72]

    In: CVPR (June 2025)

    Zielonka, W., Bolkart, T., Beeler, T., Thies, J.: Gaussian eigen models for human heads. In: CVPR (June 2025)

    2025

  73. [73]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

    Zielonka, W., Bolkart, T., Thies, J.: Instant volumetric head avatars. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4574–4584 (2023)

    2023

  74. [74]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR) (June 2025)

    Zielonka, W., Garbin, S.J., Lattas, A., Kopanas, G., Gotardo, P., Beeler, T., Thies, J., Bolkart, T.: Synthetic prior for few-shot drivable head avatar inversion. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR) (June 2025)

    2025

  75. [75]

    Which rendering looks more realistic, especially in the hair region?

    Zielonka, W., Kirschstein, T., Bolkart, T., Giebenhain, S., Sklyarova, V., Deng, X., Xiang, D., Saito, S., Liu, Y., Nießner, M., Thies, J.: How to build digital humans? from priors to photorealistic avatars. Computer Graphics Forum (Eurographics State-of-the-Art Report)45(2) (2026) Head Avatars with Dynamic Explicit Hair 21 A Background A.1 Dynamic Head A...

    2026