acceptodds
Under review as a conference paper at ICLR 2027

Photorealistic Dynamic Hair Reconstruction with Strand Gaussians

Abstract

Reconstructing dynamic human hair from multi-view videos is essential for creating high-quality 3D assets for digital avatars and animation. Existing methods augment static hair reconstruction with learned hair dynamics, facing challenges in balancing frame-wise photometric fidelity and optimization efficiency. We present a Strand Gaussian based framework for reconstructing time-varying hair geometry while maintaining photometric fidelity for the observed sequence and preserving photorealistic appearance. Qualitative and quantitative evaluations on the public NeRSemble dataset suggests that our method achieves reconstruction quality that outperforms state-of-the-art methods by a significant margin.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Reconstruction, dynamic hair reconstruction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Cw5xOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Given multiview videos, the proposed method reconstructs dynamic hair strands, including its appearance. It consists of two parts: a) coarse strand generation module and b) refinement module that optimizes for Gaussian SH color. A coarse strand generation module adds K-Plane features over scalp UV and time to features from a U-Net over the UV grid. It decodes them into strand geometry with the frozen geometry decoder of Gaussian Haircut. Training uses L1, mask, orientation, strand-smoothness, scalp-SDF and temporal-smoothness losses. The method is evaluated on 6 subjects on Nersemble dataset. It is compared to 4DGS, GaussianAvatars and Gaussian Head Avatar using PSNR/SSIM/LPIPS, with an ablation of three components.

Critical strengths
  1. Explicit, strand-based dynamic hair is useful for avatars than unstructured Gaussians. The hairstyle-transfer demo (Fig. 9) hints at reusable assets.
  2. The paper is generally easy to follow.
Critical weaknesses
  1. [Not enough qualitative results] There is not enough qualitative results. Table-1 states that metrics are averaged over six NeRSemble subjects, however there is not enough video results. The video includes only single actor.

  2. [Marginal Improvement] The abstract states the method “outperforms state-of-the-art methods by a significant margin” (L040-041). In Tab. 1 the PSNR gain over GaussianAvatars is 0.25 dB (31.59 vs 31.34)

  3. [Efficiency motivation not supported] The motivation is that other dynamic-hair methods are costly (L055-058), and the abstract claims a balance of “optimization efficiency”. No training time, memory, or rendering speed is compared with any baseline. The appendix only mentions ~12 hours for 20k strands.

  4. The “dynamic” aspect is not tested. Evaluation uses one held-out view of the training frames, so it measures view interpolation of a per sequence fit. There are no held-out time steps.

Questions for authors
  1. Fig. 1 shows GaussianAvatars at 20.92 dB, about 10 dB below its Tab. 1 average of 31.34. Could authors clarify whether this is an outlier subject?
  2. Would it be possible for authors to show the inner part of the hair?
8 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer Cw5x

Thank you for your review. We provide the following explanations to address your concerns.

Response to W1

A: Thank you for the suggestion. We have added the supplementary video comparison_2x5_4s.mp4, showing rendered appearance and 3D strand motion for five subjects, alongside the seven-subject appearance and geometry results in appendix Figs. 8–9. These results demonstrate detailed hair appearance and coherent strand motion across the presented subjects and hairstyles, providing broader qualitative support for the robustness of our reconstruction.

Response to W2

A: We agree that the average PSNR gain over GaussianAvatars is modest (0.25 dB), and the phrase “by a significant margin” was too strong. Table 1 also shows a reduction in LPIPS from 0.0896 to 0.0764 (14.7%), although our SSIM is lower. Together with the reconstruction of explicit time-varying strand geometry, these results describe our contribution more accurately. We have revised the abstract accordingly.

Response to W3

A: Thank you for this comment. Our efficiency motivation concerns motion-history encoding: methods such as DynHair use an LSTM to process a temporal window through recurrent state updates, introducing sequential computation. Our representation directly queries strand geometry and appearance from UV coordinates and time, avoiding this recurrent history-encoding step. We agree that our original wording overstated the efficiency implications and have moderated the abstract and introduction to present this as a design motivation. The approximately 12-hour training time reported in the appendix documents our computational cost.

Response to W4

A: Our objective is to recover a dynamic hair asset from a given multi-view video, including temporally corresponding strands and their changing geometry and appearance. The held-out-view evaluation assesses appearance fidelity. Fig. 7 and the supplementary video comparison_2x3.mp4 compare our method with the variant without K-Plane, showing hair deformation relative to the head, while Table 3 quantifies motion smoothness and stability. Together, these evaluations support coherent reconstruction of the observed hair motion throughout the captured sequence. More broadly, a common approach to hair-motion modeling predicts hair states from motion conditions and temporal history, without direct temporal interpolation. For example, NeuWigs evolves the previous hair state using recent head motion and gravity, while DynHair predicts deformation of a canonical hairstyle from an LSTM encoding of head-motion history.

Response to Q1

A: The 20.92 dB PSNR in Fig. 1 is computed for the single image shown in the teaser, while Table 1 reports results averaged over six subjects. As visible in Fig. 1, the hair boundary reconstructed by GaussianAvatars is misaligned with the GT boundary in this example, contributing to its pixel-wise reconstruction error. We evaluate reconstruction quality in the hair region. We hope this clarifies the difference between the score for the illustrated image and the aggregate results in Table 1.

Response to Q2

A: We have added internal hair visualizations in the final part of the appendix, “Internal Hair Structure” (Fig. 11). We apply a cutting plane to the hair on the right side of the visualization, rendering the hair above the plane normally and the hair below it semi-transparently. This reveals the internal strands while preserving the overall hairstyle structure, with an enlarged view showing their geometric details.

AuthorsOpenReview ↗

A Gentle Reminder Regarding Our Rebuttal and Revised Manuscript

Dear Reviewer,

As we approach the end of the discussion period, we hope you have had time to review our rebuttal and revised manuscript. If our responses have adequately addressed your concerns, we would sincerely appreciate it if you could consider updating your rating. Otherwise, we would be grateful for any further feedback on the remaining issues.

Thank you very much for your time and thoughtful feedback.

Reviewer_Cw5xOpenReview ↗

Concerns are addressed, quality is still limited

Dear authors, thank you for your detailed answers! Thanks for sharing the additional videos, especially comparison_2x5_4s.mp4. Some of my concerns have been addressed. I have a follow-up question: By 'inner part of the hair' I meant the hair gaussians without head to see how the internal Gaussians look like. Would it be possible to render the hair Gaussians only?

Regarding the evaluation, I'm happy to increase my score if other reviewers have a strong argument, because the quality of the hair appearance seems blurry to me.

Reviewer_NndEOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework for photorealistic dynamic hair reconstruction from multi-view videos using strand-aligned Gaussian primitives. The method consists of a strand generation module and an appearance refinement module. The former combines scalp UV-time K-Plane features with U-Net-based spatial features to predict time-varying strand geometry and appearance. The latter introduces a strand-aligned UVL-time representation to predict residual spherical harmonic coefficients for individual strand Gaussians. Geometric and temporal regularization terms are incorporated to encourage plausible strand structures and coherent motion.

Experiments on the NeRSemble dataset demonstrate improvements in average PSNR and LPIPS over the evaluated Gaussian-based baselines, although the proposed method does not achieve the best SSIM.

Critical strengths
  1. Structured representation with promising appearance refinement.

The combination of explicit hair strands and Gaussian rendering provides a useful representation for reconstructing dynamic hairstyles. In particular, the proposed strand-aligned UVL-time appearance refinement is well motivated, as it associates appearance features with the intrinsic coordinates of individual strands rather than their time-varying Cartesian positions.

  1. Encouraging perceptual reconstruction quality.

The method demonstrates visually detailed hair reconstructions across some hairstyles and achieves the lowest average LPIPS among the evaluated methods. The qualitative results suggest that the representation can preserve fine directional texture and plausible strand structure.

Critical weaknesses
  1. Incremental improvement from previous open-source project,Gaussian Haircut (Zakharov et al., ECCV 2024) .

It looks like most modules were introduced by the Gaussian Haircut, although this paper claim the dynamic hair application but no comparison to show the difference.

  1. Insufficient evidence for accurate dynamic geometry and motion reconstruction.

This paper claims it as a dynamic hair reconstruction but only show six samples for evaluation under limited hair styles. In the supplementary , basically the hair styles move with head rigidly.

  1. Incomplete justification of the proposed components and optimization objectives.

The ablation results reveal an important trade-off: removing either the SDF loss or the hair smoothness loss improves all three reported image-based metrics, while the paper primarily relies on qualitative examples to justify these regularizers. Quantitative geometric measurements are needed to establish their benefits. The ablation study did not show the significance of these components.

  1. Missing reference, Difflocks(CVPR2025), MonoHair(CVPR2024), EfficientMonoHair(CVPR2026), dynamic gaussian hair(NeurIPS2025)

  2. Error in Eq.5 about the opacity with a missing exp function.

Questions for authors
  1. Comparison to prior strand-based methods. What is the principal methodological advantage over a temporal extension of Gaussian Haircut? Can you use the synthetic dataset to get the geometry metrics like the dynamic gaussian hair(NeurIPS2025)?

  2. Component contributions. Could the authors report ablations isolating the K-Plane and U-Net branches and the coarse-only reconstruction? This would clarify whether the gains arise primarily from temporal feature modeling, refinement capacity, or the UVL coordinate parameterization.

  3. Temporal loss and coordinate system. In which coordinate system is Eq. (18) evaluated? If it is applied to world-space strand points, how does the method avoid suppressing legitimate rigid head motion? Additionally, could the authors confirm the Gaussian opacity formulation in Eq. (5)?

  4. Could author explain the ablation study metrics decrease when you remove the loss in Table2(Line 464) and how to explain the visualization improvemnet(although it is trivial). Is it possible to use some geometry metrics(Like the orientation volume accuracy used in the strand reconstruction) to show the benefit of these losses?

5 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer NndE (1/2)

Thank you for your review. We provide the following explanations to address your concerns.

Response to W1

A: Our method builds on Gaussian Haircut’s strand-aligned Gaussian representation. Our contribution is the joint reconstruction of observed dynamic hair sequences, with two main differences:

  1. Joint reconstruction across time. Gaussian Haircut reconstructs a single hairstyle. We optimize a shared scalp UV–time representation that decodes changing strand geometry and appearance throughout the captured sequence. Observations across frames jointly constrain this representation. U-Net provides a spatial prior to stabilize optimization.
  2. Appearance refinement on moving strands. Gaussian Haircut directly refines decoded strand coordinates in its fine stage. Our refinement preserves the reconstructed dynamic geometry and predicts spherical-harmonic residuals in (u,v,ℓ,t)(u,v,\ell,t) coordinates. This associates appearance corrections with moving strands and recovers details smoothed by guiding-strand interpolation.

Fig. 7 and Table 4 examine the K-Plane branch and coarse-only reconstruction, while Table 2 evaluates strand-aligned UVL coordinates against XYZ coordinates. These ablations support the roles of temporal modeling and dynamic appearance refinement.

Response to W2

A: To address this concern, we clarify the coverage of our evaluation and use the K-Plane ablation (Fig. 7) and supplementary video comparison_2x3.mp4 to demonstrate reconstructed hair deformation relative to the head.

  1. Evaluation subjects and hairstyles. Table 1 reports quantitative results for six subjects, while the appendix presents qualitative reconstructions for seven subjects. These examples include short and shoulder-length hairstyles, straight and wavy hair, and tied-back hair, covering differences in length, strand flow, and overall volume. Figs. 8–9 show both appearance and reconstructed strand geometry, and we have added the supplementary video comparison_2x5_4s.mp4 to show reconstructed motion across additional subjects. Together, these results provide broader evidence of reconstruction quality and stability across the demonstrated hairstyles.
  2. Evidence from the temporal-modeling ablation. Fig. 7 and the supplementary video comparison_2x3.mp4 compare our method with the variant without the K-Plane branch. This variant retains the time-independent U-Net spatial features, so its hair largely follows the head motion with limited relative deformation, closer to the rigid behavior described by the reviewer. In contrast, the full model reconstructs changing strand configurations relative to the head, with a clear difference visible in the video. K-Plane conditions strand generation on time, allowing us to reconstruct the observed hair motion. This comparison directly illustrates the contribution of temporal modeling to non-rigid hair reconstruction.
Response to W3

A: These regularizers target geometric plausibility and temporal stability. Their benefits are therefore assessed through geometry visualizations and additional quantitative motion metrics.

  1. SDF loss. The SDF loss alleviates hair penetration into the scalp, as illustrated in Fig. 4.
  2. Hair and temporal smoothness losses. We have added quantitative evaluations following DynHair: Head Avatars with Dynamic Explicit Hair, using VER, P95-Disp, CTS, and AMD to assess extreme speeds, displacement spikes, curvature jitter, and angular momentum drift, respectively. Table 3 shows that hair smoothness reduces CTS from 0.006764 to 0.004293 (36.5%) and also improves VER and P95-Disp, with a small 1.3% increase in AMD. The lower CTS indicates less curvature jitter across frames and more stable strand geometry. Temporal smoothness improves all four metrics, including reducing P95-Disp from 0.125644 to 0.063550. These results support reduced displacement spikes and more stable reconstructed motion. Together with the visual ablations, these results provide quantitative evidence that the losses improve the smoothness and temporal stability of reconstructed strand geometry.

Table 3. Hair and temporal smoothness ablations on subject 304. Lower values are preferred.

VariantVER ↓P95-Disp ↓CTS ↓AMD ↓
w/o hair smoothness loss9.87120.0658400.0067641489.92
w/o time smoothness loss15.10600.1256440.0058101773.85
Ours9.08400.0635500.0042931509.81
Response to W4

A: Thank you for the suggestion. We have added the suggested references to the Related Work section.

Response to W5

A: Thank you for pointing this out. We have corrected Eq. (5) in Section 3 (Preliminary).

AuthorsOpenReview ↗

First Rebuttal for Reviewer NndE (2/2)

Response to Q1

A: Our responses to W1 and W2 jointly address the methodological advantage and its experimental evidence. As explained in W1, we jointly reconstruct the observed dynamic strand sequence through a shared scalp UV–time representation and refine appearance in (u,v,ℓ,t)(u,v,\ell,t) coordinates, keeping appearance corrections attached to moving strands. W2 provides direct evidence through Fig. 7 and the supplementary video comparison_2x3.mp4: removing K-Plane produces hair motion closer to rigid head motion, while the full model reconstructs changing strand configurations relative to the head.

For quantitative geometric evaluation, we have added the four motion metrics in Table 3, as discussed in W3. These assess strand-motion stability and smoothness. Evaluation against synthetic ground-truth strand geometry remains to be conducted.

Response to Q2

A: U-Net provides time-independent spatial features shared across frames, giving strand optimization a stable spatial foundation. Its role is similar in motivation to the initial static reconstruction in 4DGS before enabling temporal deformation. We have also added K-Plane removal and coarse-only ablations in Fig. 7 and Table 4, showing that temporal modeling and appearance refinement improve reconstruction quality.

Table 4. K-Plane branch and refinement ablations. All results are averaged over all frames of subject 304. The first two rows show coarse-stage results.

VariantPSNR ↑SSIM ↑LPIPS ↓
w/o K-Plane27.440.88640.1042
Coarse31.730.90050.0770
Ours32.220.90880.0562
Response to Q3

A: Coordinate system and temporal regularization. The time smoothness loss is computed on reconstructed strand points in world coordinates. It encourages temporal coherence of the reconstructed strand motion. Head motion is prescribed by the tracked FLAME meshes, with strand roots attached to the scalp. The temporal term is a soft regularizer balanced against multi-view image, mask, and orientation supervision. It can also penalize accelerated rigid motion, including rotation, so its weight controls the trade-off between smoothness and fitting the observed motion. Fig. 6 and Table 3 show its benefits for temporal coherence and strand-motion stability.

Opacity formulation. Thank you for pointing this out. We have corrected Eq. (5) in Section 3 (Preliminary).

Response to Q4

A: Removing geometric regularization allows more freedom in image fitting, which can improve image metrics while introducing geometric artifacts. As detailed in our response to W3, we have added quantitative evaluations in Table 3 that support improved strand smoothness and motion stability with regularization.

AuthorsOpenReview ↗

A Gentle Reminder Regarding Our Rebuttal and Revised Manuscript

Dear Reviewer,

As we approach the end of the discussion period, we hope you have had time to review our rebuttal and revised manuscript. If our responses have adequately addressed your concerns, we would sincerely appreciate it if you could consider updating your rating. Otherwise, we would be grateful for any further feedback on the remaining issues.

Thank you very much for your time and thoughtful feedback.

Reviewer_oXCfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for reconstructing dynamic hair from multi-view videos using strand Gaussians. The strand generation module combines scalp UV-space features with time-aware K-Planes features to generate guiding strands, then constructs dense strands through interpolation. The strand generation module is subsequently frozen, and color residuals are predicted to further refine appearance. Experiments demonstrate improved metrics over methods such as GaussianAvatars.

Critical strengths

S1. The method adopts a well-structured coarse-to-fine reconstruction framework with spatial and temporal regularization to support plausible strand geometry and coherent hair motion.

S2. The paper is clearly written, logically organized, and precise in its presentation.

Critical weaknesses

W1. DGH (Wang et al., 2025) already combines dynamic hair modeling with strand-Gaussian appearance optimization, FHAvatar (Sun et al., 2026) decodes strand Gaussians in UV space, and Perm (He et al., 2025) introduces hierarchical representations separating global hair structure from local strand details. Against these precedents, the proposed UV–time features and strand-aligned color residuals appear to constitute an incremental contribution.

W2. The supplementary video shows visible gaps between the face (scalp) and hair during large motions, raising concerns about their dynamic alignment. The lack of analysis of these artifacts limits confidence in the method’s robustness under such motions.

Questions for authors

Please discuss and compare the proposed method with the aforementioned works and other recent hair reconstruction methods, explaining what technical contributions and benefits extend beyond these precedents.

6 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer oXCf

Thank you for your review. We provide the following explanations to address your concerns.

Response to W1/Q1

A: We appreciate the reviewer highlighting these closely related works. Their objectives and experimental settings differ in several respects from ours, which focuses on extracting dynamic hair geometry and appearance from multi-view videos:

  1. DGH learns hair dynamics from simulated data to animate an existing hairstyle, while we reconstruct the motion observed in captured videos.
  2. FHAvatar first reconstructs a static canonical face-and-hair model from a few unordered images for subsequent FLAME-driven animation, while we recover time-varying strand geometry and appearance from video.
  3. Perm separates global structure and local curl patterns of static hairstyles, while our coarse and refinement stages recover dynamic geometry and refine appearance.

Our technical contribution addresses two coupled reconstruction challenges within a consistent strand parameterization:

  1. Joint reconstruction of dynamic strand sequences. A shared set of scalp UV roots identifies strands across frames. We jointly optimize a scalp UV-time representation over the input sequence to decode time-varying strand geometry and appearance. Observations across frames constrain the same representation, linking the reconstructed states through a common strand structure. K-Plane models temporal variation, while U-Net provides a spatial prior to stabilize optimization.
  2. Appearance refinement attached to moving strands. To recover details smoothed by guiding-strand interpolation, we predict per-Gaussian spherical-harmonic residuals in (u,v,ℓ,t)(u,v,\ell,t) while keeping the reconstructed geometry fixed. The root coordinates identify a strand, and ℓ\ell identifies a position along it, so each correction remains associated with the moving strand segment. Fig. 7 illustrates the appearance improvement after refinement. In Table 2, this parameterization reduces LPIPS from 0.0658 with XYZ coordinates to 0.0562.
Response to W2

A: To clarify reconstruction stability under large head motions, we explain scalp attachment, hair-region supervision, and the additional reconstruction evidence:

  1. Stable scalp attachment. Strand roots at the hairline and temples are bound to the tracked FLAME scalp at each frame, so their positions follow the scalp throughout head motion.
  2. Hair alignment and video compositing. The mask loss in Eq. (19) supervises alignment with the segmented hair region, where we also evaluate reconstruction quality. In the supplementary videos comparison_2x5_4s.mp4 and comparison_2x3.mp4, we composite GT facial pixels with rendered hair using a face mask. Imperfect matching between the face and hair masks can leave visible gaps at their boundary. This compositing step does not change the reconstructed strand geometry or scalp anchors, so these display gaps do not by themselves indicate a loss of scalp attachment. Segmentation errors remain an input limitation, as discussed in Sec. 6.
  3. Evidence across subjects. Appendix Figs. 8–9 and comparison_2x5_4s.mp4 show reconstructed appearance and 3D strand motion across additional subjects and hairstyles, providing qualitative support for reconstruction stability in the demonstrated sequences.
AuthorsOpenReview ↗

A Gentle Reminder Regarding Our Rebuttal and Revised Manuscript

Dear Reviewer,

As we approach the end of the discussion period, we hope you have had time to review our rebuttal and revised manuscript. If our responses have adequately addressed your concerns, we would sincerely appreciate it if you could consider updating your rating. Otherwise, we would be grateful for any further feedback on the remaining issues.

Thank you very much for your time and thoughtful feedback.

I appreciate the authors' updates, additional video results, and ablation studies. Some of my concerns have been addressed.

However, I will maintain my original score for now, as the video results remain limited, making it difficult to assess potential blurring of non-surface hair strands and at extreme viewing angles. Additionally, the discussion of the missing reference in W1 is not fully reflected in the paper.

I am open to further discussion or raising my score if other reviewers feel strongly otherwise.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their constructive comments and suggestions. We have carefully considered each point and uploaded revised versions of the manuscript and supplementary material. Changes in the manuscript are marked in blue. We have also added two supplementary videos:

  1. comparison_2x3.mp4 presents the K-Plane ablation in a two-row, three-column layout. From left to right, the columns show GT, coarse-stage reconstruction with K-Plane, and coarse-stage reconstruction without K-Plane. The top row shows appearance, and the bottom row shows strand geometry, with the GT image repeated in the bottom-left panel for reference. This comparison illustrates hair deformation relative to the head.
  2. comparison_2x5_4s.mp4 presents additional reconstruction results in a two-row, five-column layout. Each column corresponds to a different subject. The top row shows rendered appearance, and the bottom row shows the corresponding strand geometry, illustrating reconstruction quality and motion across different subjects and hairstyles.

The main revisions are as follows:

  1. Expanded the Related Work section with the suggested references (Sec. 2).
  2. Corrected the missing exponential in the Gaussian opacity formulation (Eq. (5), Sec. 3).
  3. Added four motion metrics, VER, P95-Disp, CTS, and AMD, in Sec. 5.1, with quantitative hair and temporal smoothness ablations in Table 3.
  4. Added coarse-stage K-Plane removal and coarse-only ablations, with visual comparisons in Fig. 7 and quantitative results in Table 4 (Sec. 5.3).
  5. Added an internal hair structure visualization in the final part of the appendix (Fig. 11).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.