acceptodds
Under review as a conference paper at ICLR 2027

StrainAvatar: Strain-Driven Head Avatars with Explicit Facial Skin Deformation

Abstract

Head avatars reconstructed from multi-view video achieve photorealistic rendering, yet expression-dependent wrinkles are typically synthesized in appearance or modeled as directly predicted displacements, leaving their geometric cause unmodeled. We introduce StrainAvatar, a head avatar that reconstructs wrinkle-scale facial deformation as explicit geometry from multi-view RGB alone. Our key idea is to factor wrinkle formation into a shared mechanical cause and an identity-specific geometric response. Facial compression, analytically characterized by the strain of the tracked deformation, provides an identity-agnostic signal that generalizes across expressions, while a subject-specific network learns only how the skin responds to this signal. We realize this factorization by predicting per-triangle Jacobian correctives from the strain field and integrating them into a coherent corrected mesh through a Poisson solve. Because a zero corrective exactly recovers the tracked mesh, large-scale facial motion remains deterministic and learning is restricted to wrinkle-scale deformation. We further transport Gaussian primitives using the deformation gradient of the corrected mesh, making appearance explicitly inherit the recovered surface deformation. Experiments show rendering quality comparable to recent head avatars, improved surface-normal accuracy, and consistent reductions in mesh error against independent multi-view stereo depth, qualitatively recovering expression-dependent wrinkle structure as explicit, animatable geometry.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Mesh representation, 3DGS, Neural Jacobian Field

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Xn2dOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents StrainAvatar, a person-specific head avatar that reconstructs expression-dependent mesh wrinkles from multi-view RGB. It uses analytically computed strain to predict residual Jacobian corrections, integrates them through a Poisson solve, and transports Gaussians using the corrected mesh deformation with surface alignment. Experiments on nine NeRSemble subjects evaluate novel views and expressions, reporting competitive rendering quality, improved normal agreement, and reduced mesh depth error.

Critical strengths
  1. Strain-driven wrinkle modeling. Using surface strain to drive a person-specific wrinkle response is a reasonable design. It gives the network a meaningful local deformation signal for learning how facial wrinkles change with expression.
  2. Image-based geometry supervision and convincing mesh results. White-texture shading and stopped lighting gradients direct the wrinkle-band loss toward mesh deformation, which then drives Gaussian rendering. The qualitative mesh comparisons show clearer and finer wrinkle structures, particularly around the forehead and between the eyebrows. The reduced stereo-depth error also supports improved surface accuracy at a coarser facial-shape scale.
Critical weaknesses

The contribution of the proposed wrinkle modeling is not fully isolated. Removing the corrected mesh eliminates all learned geometric refinement, which may include personalized shape correction as well as wrinkle detail. The targeted geometry-stage ablations in Fig. 7, without the strain map or wrinkle-band loss, would help clarify this attribution, but are reported only qualitatively. Table 4 instead tests strain conditioning in the Gaussian decoders, a different component. Section 4.3 also states that the Fig. 7 variants increase stereo-depth error and refers to Appendix H, yet the corresponding numbers are not reported there. Since the depth and pseudo-normal references do not resolve individual wrinkles, quantitative results for these targeted ablations are particularly important for assessing the central contribution.

Questions for authors
  1. Quantitative results for the geometry ablations: Could the authors provide quantitative geometry and final rendering results for the variants without the strain map or wrinkle-band loss, while retaining geometry optimization and matched subsequent Gaussian training? Section 4.3 refers to increased stereo-depth errors in Appendix H, but these measurements do not appear to be listed there.
  2. Choice of deformation representation: Could the authors elaborate on the benefits of predicting Jacobian corrections followed by Poisson integration, compared with directly predicting vertex offsets or a UV displacement map? If available, a comparison using the same strain inputs and geometry supervision would help clarify the contribution of this design choice.
5 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for recognizing the motivation of strain-driven wrinkle modeling and the qualitative geometric improvements. Below, we address the geometry-stage ablations and the choice of deformation representation.


W1. Isolating the Geometry-Stage Contribution

We acknowledge this oversight and thank the reviewer for pointing it out. The geometry-stage measurements referenced in Sec. 4.3 were omitted from Appendix H. We report them in Q1 below, together with final rendering results under matched appearance-stage training. These ablations remove the strain map T\mathbf{T} from the corrective network Φ\Phi or disable the wrinkle-band loss Lwb\mathcal{L}_{\text{wb}}, while retaining geometry optimization. They assess the geometry stage separately from the Gaussian-decoder strain ablation in Table 4.


Q1. Quantitative Geometry and Rendering Ablations

For each variant, we retain geometry-stage optimization and either set the input strain map T\mathbf{T} of the corrective network Φ\Phi to zero or disable Lwb\mathcal{L}_{\text{wb}}. All other geometry-stage losses and the 200k-iteration budget remain the same. We then retrain the appearance stage on each resulting corrected mesh Vw\mathbf{V}_w for 600k iterations, using the same hyperparameters and FLAME finetuning protocol.

On subject 104 under Novel-View Synthesis, we evaluate the MAE of the rasterized mesh depth against the same stereo-depth reference and mask used in Appendix H.

SurfaceDepth MAE ↓ (mm)
Tracked mesh Vc\mathbf{V}_c2.26
Corrected mesh Vw\mathbf{V}_w (ours)1.69
−- strain map T\mathbf{T} to Φ\Phi1.99
−- wrinkle-band loss Lwb\mathcal{L}_{\text{wb}}2.10

For the final appearance-stage outputs, we report rendering quality and Gaussian-normal NCS against DAViD over the full foreground, following Table 4:

VariantPSNR ↑SSIM ↑LPIPS ↓NCS ↑
Ours34.460.9460.0750.912
−- strain map T\mathbf{T} to Φ\Phi34.480.9450.0760.909
−- wrinkle-band loss Lwb\mathcal{L}_{\text{wb}}34.260.9460.0750.909

Both ablations increase mesh depth MAE, while the differences in rendering quality remain small. This is consistent with the appearance stage absorbing the photometric residual, as discussed in Sec. 4.3; image similarity alone does not isolate the contribution of the recovered geometry.


Q2. Choice of Deformation Representation

The main benefit is that the network predicts local Jacobian correctives ΔJf\Delta\mathbf{J}^f, and the Poisson solve integrates the corrected Jacobian field into a single coherent surface Vw\mathbf{V}_w. Compared with directly adding predicted displacements, this jointly determines vertex positions by fitting the local deformation gradients under the tracked face-boundary constraints.

We compare this formulation with direct UV displacement prediction using the same strain map T\mathbf{T}, U-Net backbone, RGB supervision, geometry-stage losses, boundary constraints, and 200k-iteration budget.

We report mesh depth MAE and mesh-normal NCS on the held-out FREE sequence of subject 210 against COLMAP MVS depth and its derived normals. We use the same hair-free masking protocol as in the paper, with frames sampled at one-second intervals.

RepresentationDepth MAE ↓ (mm)MVS NCS ↑
Strain →\rightarrow UV displacement1.6280.9566
Strain →ΔJ\rightarrow\Delta\mathbf{J} + Poisson (ours)1.4980.9604

Jacobian correctives followed by canonical Poisson integration reduce mesh depth MAE by approximately 8.0% under matched RGB supervision, supporting this design choice in the evaluated setting.

We have uploaded a preliminary revised manuscript with additional experiments and clarifications highlighted in blue, and will continue to update it during the discussion period.

Thank you for the additional experiments. Since the final rendering PSNR gains from strain conditioning and the wrinkle-band loss appear marginal, my assessment relies primarily on their geometric benefits. However, the Q1 geometry-stage ablations cover only subject 104 under novel-view synthesis, leaving their consistency across identities unclear. Could you repeat these two ablations on one additional original subject using the same evaluation protocol and matched geometry-training settings? Geometry metrics alone would suffice; appearance-stage retraining is unnecessary. Consistent improvements would strengthen the case for raising my rating to weak accept.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for this suggestion. We repeated both geometry-stage ablations on all subjects using the same evaluation settings as in Q1. We report mesh depth MAE (mm) as mean ± std across subjects.

SurfaceNovel View MAE ↓Novel Expression MAE ↓
Tracked mesh1.89 ± 0.432.97 ± 0.86
Corrected mesh (ours)1.17 ± 0.282.46 ± 0.72
− strain map1.47 ± 0.292.63 ± 0.75
− wrinkle-band loss1.84 ± 0.442.99 ± 0.82

On average, removing either the strain input to the corrective network or the wrinkle-band loss increases mesh depth MAE on both splits. These results extend the evidence for their geometric contributions beyond subject 104.

Reviewer_4MNcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes StrainAvatar, per-subject, multi-view head avatar that bake wrinkles into the mesh geometry. The Green-Lagrange strain of the tracked FLAME deformation is fed to a U-Net to predict per-triangle Jacobian correctives, with Poisson solve mesh correction. Gaussians are carried by the corrected mesh’s deformation gradient. Render quality is comparable to TexAvatars on 9 NeRSemble identities with best normal similarity to DAViD pseudo GT.

Critical strengths
  1. Jacobian correctives prediction is well-grounded via Green-Lagrange strain such that it bounds to wrinkle geometry which is nontrivial in head avatar photorealism. Zero corrective reproduces the tracker exactly. The input strain field is local and invariant to rigid transformation, grounded by traditional graphics theory. Disentangling appearance and geometry for wrinkle contributes to photorealism with nuanced expression. I appreciate this strength since it’s a standalone problem definition and solution pair.
Critical weaknesses
  1. The effect of strain-driven geometry is not sufficiently analyzed and reported. Removing the corrected mesh leaves NCS unchanged (0.912, Table 4) while adding B or L_align improves it. The normal gains therefore might come from keeping Gaussians flat on the surface, not from the recovered geometry since DAViD normals cannot resolve wrinkles (line 471). Figure 7. Includes strain map ablation but no corresponding quantitative results. Normal similarity has been measured on pseudo GT and reported numbers lack statistical evidence (e.g. multi-seed runs with std and various subjects). I significantly take these factors into consideration for my rating.

  2. Limited experiments. The contribution is restricted to wrinkles in that marginal novel expression gain over TexAvatars even mostly comes from subject 253 and 460 while on the other seven it is negligible. NVS PSNR is lower on 7 of 9 subjects compared to TexAvatars. Wrinkles are mostly prominent in unseen extreme expressions but authors do not report ablation on novel & FREE(in TexAvatars protocol) expression and quantitative comparisons with baselines on FREE expression. I somewhat care these factors into consideration for rating.

Questions for authors
  1. Could authors provide alternative measurement of wrinkle-level normal if current NCS metric is unreliable for such details?
  2. Additional quantitative results including MeGA under it’s masked region?
  3. Could more details (e.g. the number of Gaussians sampled from UV map, FPS comparisons, etc.) and additional experiments in W2 be provided for proving statistical significance and generalization?

I would like to increase my score if all weaknesses and questions are well explained / addressed.

8 comments
AuthorsOpenReview ↗

Official Comment by Authors (1)

We thank the reviewer for recognizing the motivation of strain-driven geometric refinement and the value of explicit wrinkle modeling. Below, we address the feedback on geometric validation, evaluation, and implementation details.


W1. Geometry Contribution, Alignment, and Statistical Evidence

Table 4 measures rendered Gaussian-normal agreement with DAViD predictions. We additionally assess the geometric contribution through direct mesh depth evaluation.

For the Table 4 ablations, we report DAViD NCS (mean ± std) under Novel-View Synthesis, computed across six runs covering three subjects (074, 104, and 253) with two seeds each. The standard deviation reflects variation across both subjects and seeds.

VariantNCS ↑
Ours0.927 ± 0.015
−- corrected mesh Vw\mathbf{V}_w0.925 ± 0.013
−- T\mathbf{T} in decoders0.913 ± 0.009
−- alignment basis B\mathbf{B}0.907 ± 0.016
−- alignment loss Lalign\mathcal{L}_{\mathrm{align}}0.882 ± 0.018

Direct surface statistics for the final Gaussian avatars across nine subjects are reported in Q1 below.

We additionally report quantitative results for the geometry-stage ablations in Fig. 7. On subject 104 under Novel-View Synthesis, we measure mesh depth MAE against the same stereo-depth reference and mask used in Appendix H.

SurfaceDepth MAE ↓ (mm)
Tracked mesh Vc\mathbf{V}_c2.26
Corrected mesh Vw\mathbf{V}_w (ours)1.69
−- strain map T\mathbf{T} to Φ\Phi1.99
−- wrinkle-band loss Lwb\mathcal{L}_{\text{wb}}2.10

Removing either component increases mesh depth MAE, supporting its contribution to mesh refinement. These geometry-stage ablations differ from removing strain conditioning in the Gaussian decoders in Table 4.


W2. Novel-Expression and FREE-Sequence Evaluation

We additionally compare all four methods for novel-expression synthesis on the held-out FREE sequences of nine subjects, using all timesteps from the held-out camera and the same FLAME face mask. This focuses the image metrics on the facial region relevant to wrinkle modeling and reduces the influence of errors in the torso and other non-facial regions.

MethodPSNR ↑SSIM ↑LPIPS ↓
GaussianAvatars30.920.9120.154
SurFhead30.600.9140.169
TexAvatars31.420.9080.084
Ours31.870.9090.082
AuthorsOpenReview ↗

Official Comment by Authors (2)

Q1. Direct Surface Evaluation against a High-Resolution MVS Reference

As the reviewer notes, DAViD normals do not resolve individual wrinkles. We therefore evaluate against a high-resolution MVS reference reconstructed from full-resolution images (2200×32082200\times3208) using all 16 calibrated cameras. This reference is used only for evaluation; our model is trained from RGB alone.

We report symmetric Chamfer distance (CD) and surface NCS over the FLAME face region of nine subjects. The table shows mean ± std across subjects, with each subject score averaged over frames sampled at one-second intervals. Both Gaussian methods use the same surface extraction.

SurfaceCD ↓ (mm)Surface [email protected] mm ↑
Tracked FLAME mesh1.929 ± 0.3130.902 ± 0.022
Corrected mesh (ours)1.760 ± 0.2530.903 ± 0.021
TexAvatars (extracted surface)1.323 ± 0.2200.900 ± 0.029
Ours (extracted surface)1.220 ± 0.2230.912 ± 0.025

The corrected mesh reduces CD independently of Gaussian alignment, while our extracted avatar surface achieves lower CD and higher surface NCS than TexAvatars.


Q2. MeGA Comparison under a Shared Mask

We compare all five methods from the held-out camera using the same MeGA evaluation mask, obtained by multiplying the parsing-based object mask by the MeGA soft rasterized full mask, including hair. Following the MeGA protocol, we mask both the rendered and ground-truth images.

The table reports mean per-subject scores over eight subjects under Novel-View Synthesis and Novel Expression.

MethodNovel-View Synthesis / PSNR ↑Novel-View Synthesis / SSIM ↑Novel-View Synthesis / LPIPS ↓Novel Expression / PSNR ↑Novel Expression / SSIM ↑Novel Expression / LPIPS ↓
GaussianAvatars35.220.9610.07332.600.9490.079
SurFhead34.590.9590.08532.520.9500.088
TexAvatars39.090.9720.02832.980.9470.048
MeGA30.670.9300.07929.930.9270.082
Ours39.010.9720.02733.510.9490.046

Under this shared mask, our method remains competitive with TexAvatars and achieves higher PSNR and SSIM and lower LPIPS than MeGA on both splits.


Q3. Implementation Details and Runtime

Gaussians are sampled from valid texels of the 5122512^2 FLAME UV map, including the tongue, with a fixed count across subjects; the corrected facial mesh has 27.6k vertices. We report FPS measured at 550×802550\times802 on a shared NVIDIA RTX 3090.

MethodPrimitive countFull pipeline FPS
Ours240.8k18
TexAvatars238.2k51
GaussianAvatars104.2k132
SurFhead64.5k surfels11
MeGA81.5k + facial mesh (16.4k vertices)7

In this benchmark, our full pipeline is faster than SurFhead and MeGA, which also model surface geometry (18 FPS versus 11 and 7 FPS, respectively). Our method reaches 35 FPS with precomputed geometry. The full pipeline includes recomputing geometry for each expression, while the canonical Poisson factorization is constructed once and reused.

The additional experiments requested in W2 are discussed above.

We have uploaded a preliminary revised manuscript with additional experiments and clarifications highlighted in blue, and will continue to update it during the discussion period.

Reviewer_4MNcOpenReview ↗

Reviewer 4MNc - Response to First Rebuttal

I appreciate the authors’ response and the additional experiments. In order to properly figure out the strenghts of StrainAvatar, I would like to check a couple more things.


W1: Thanks for providing multi-seed runs which improve the reliability of the results. Now I could check some components have actual impact on the geometry.

However, still geometry-stage ablations (Table 6. in the revised PDF) are provided with a single run while they are the most significant in claiming lower error in strain-related geometry of StrainAvatar. Also, the negligible improvement in normals by adding the corrected mesh is not justified without discussion. It improves coarse geometry (Chamfer in Tab. 4., depth in Tab. 6. in the revised PDF), but not normals, where wrinkles mainly appear and the submission mostly illustrates as qualitative evidences, so the wrinkle-scale benefit is still not shown.

W2: Introducing FREE corpus comparison with baselines (Tab. 15) now supports the advantage under unseen expressions. However, I am still worried that the novel-expression gain over TexAvatars mostly comes from subjects 253 and 460 and is negligible on the other seven, which is not analyzed. Readers would love to be suggested with such discussion.

Q1: Thank you for providing the high-resolution MVS evaluation over nine subjects, which addresses my question. Would you please provide the exact definition of [email protected] (mm) metric although I can assume, for clarity? Additionally, why MeGA is not included in this table?

Q2 & Q3: Thanks for addressing my concerns about missing quantitative comparison of MeGA and implementation details.

Overall, marginal gains in the rebuttal, such as the single-run geometry ablations (W1) and Tab. 7 on one subject, need to be more justified with statistically significant evidence. If you think there are certain reasons, would you please explain? I am aware that StrainAvatar’s contribution is better shown in qualitative results rather than numerical figures because of its own nature.

Given the current status, I recommend the authors provide additional experiments with enough analyses and discussions on these points, which could justify the claims given the relatively slow FPS (in rebuttal to Q3). I acknowledge that the core idea is interesting and mathematically well-bounded, so I look forward to hearing back from you.

Best Regards,

Reviewer 4MNc

Reviewer_aBiwOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

StrainAvatar reconstructs expression-dependent wrinkles as explicit mesh geometry from multi-view RGB images. It computes strain from tracked FLAME meshes, predicts identity-specific per-triangle Jacobian correctives, and integrates them through a Poisson solve. The corrected mesh then drives the deformation of 3D Gaussians. Experiments demonstrate improved facial geometry and qualitative wrinkle recovery.

Critical strengths
  1. Strain-conditioned facial deformation is well motivated. Explicit, expression-dependent wrinkle geometry offers value beyond rendering quality.

  2. The residual Jacobian formulation provides practical benefits, including recovery of the tracked mesh at zero correction, boundary continuity, and a prefactorized Poisson system.

  3. Tables 3 and 5 show consistent improvements in mesh depth accuracy, while qualitative results demonstrate additional facial details.

Critical weaknesses
  1. The quantitative evaluations do not directly validate wrinkle-scale geometry. Table 2 evaluates rendered Gaussian normals rather than the corrected mesh, while stereo depth mainly measures coarse facial shape. Neither reference resolves individual wrinkles, as the authors acknowledge.

  2. The ablations do not sufficiently isolate the benefits of strain conditioning and Jacobian prediction. Since strain is the corrective network’s only input, removing it also removes expression-dependent conditioning. Comparisons with alternative conditioning signals and displacement prediction are needed.

  3. The wrinkle-band loss may confuse geometric detail with appearance. High-frequency RGB structure can also arise from pigmentation, facial hair, specular highlights, or shadows. Its robustness to these factors remains unclear.

  4. Table 4 shows limited or mixed improvements across metrics. Removing the corrected mesh leaves NCS unchanged, while removing the alignment loss slightly improves LPIPS despite reducing NCS.

  5. The supplementary video shows changing expressions from an apparently fixed held-out camera; a camera sweep would better demonstrate view consistency. The normal maps also appear to contain artifacts around the open mouth and bangs (e.g., 0:09 and 1:21), whose source remains unclear.

Questions for authors
  1. Under matched supervision and training budgets, how does strain-conditioned Jacobian prediction compare with expression-conditioned Jacobian prediction and strain-conditioned displacement prediction?

  2. Can the authors directly evaluate fine-scale wrinkle geometry, beyond coarse depth errors and rendered Gaussian normal similarity?

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1)

We thank the reviewer for recognizing the motivation of strain-driven wrinkle modeling and the practical benefits of our residual Jacobian formulation. Below, we address the constructive feedback on geometric validation, ablations, and visualizations. We will update this response with additional experimental results as they become available.

W1. Direct Evaluation of Wrinkle-Scale Geometry

The stereo depth used in our original evaluation mainly captures coarse facial shape and provides limited information about individual wrinkles. To evaluate finer geometric detail directly, we constructed a higher-resolution MVS reference from full-resolution images (2200×32082200\times3208) using all 16 calibrated cameras. We ran COLMAP PatchMatch with fixed calibration, followed by Poisson surface reconstruction. All meshes are compared directly in the shared calibrated coordinate system, without per-mesh ICP alignment.

We evaluated the held-out FREE sequence of subject 210, which is excluded from training.

We evaluate the tracked and corrected meshes over the FLAME face region, without involving Gaussians. We report symmetric Chamfer distance (CD) and surface normal cosine similarity (NCS) against the MVS reference.

SurfaceCD ↓ (mm)Surface NCS ↑
Tracked FLAME mesh1.5430.927
Corrected mesh1.3750.931

As shown in the table, the corrected mesh achieves lower CD and higher surface NCS than the tracked mesh on the held-out FREE sequence, independently of Gaussian alignment.

W2 / Q1. Conditioning and Deformation Representation

We compare strain-conditioned Jacobian correctives with three alternatives: Jacobian correctives conditioned on a UV map of FLAME expression-and-jaw offsets (Expr$\rightarrow\Delta\mathbf{J}),strain−conditioneddirectUVdisplacements(Strain), strain-conditioned direct UV displacements (Strain\rightarrow$disp), and expression-code-conditioned displacements (Expr$\rightarrow$disp). All variants use the same RGB supervision, regularization, and 200k-iteration geometry-stage budget.

We report normal cosine similarity (NCS) of rasterized mesh normals over the FLAME face region against DAViD predictions, without involving Gaussians. Additionally, we measure NCS against depth-derived normals obtained using the MVS reconstruction protocol described in W1, with frames sampled at fixed one-second intervals for the MVS evaluation. We evaluate subject 210 under Novel Expression and on the held-out FREE sequence.

Mesh-normal NCS ↑

VariantDAViD / Novel ExpressionDAViD / FREEMVS normals / Novel ExpressionMVS normals / FREE
Strain$\rightarrow\Delta\mathbf{J}$ (ours)0.96790.96890.96030.9602
Expr$\rightarrow\Delta\mathbf{J}$0.96700.96790.95940.9591
Strain$\rightarrow$disp0.96530.96770.95510.9563
Expr$\rightarrow$disp0.96560.96680.95880.9574

W3. Geometry–Appearance Ambiguity in the Wrinkle-Band Loss

We agree that high-frequency RGB structure is not uniquely attributable to geometry. Our goal is to recover expression-dependent, wrinkle-scale facial deformation from RGB observations, using skin strain as the conditioning signal. The wrinkle-band loss compares high-pass log-luminance structure between the input and a diffuse-shading rendering with a uniform white texture. Gradients to lighting are stopped, so this term updates geometry through surface normals.

The RGB reconstruction term and auxiliary texture and lighting models help explain appearance variations, while local normalized cross-correlation (NCC) reduces sensitivity to intensity gain and bias. Multi-view supervision and deformation regularization further constrain the fitting.

Fig. 6 shows recovered wrinkle geometry across multiple subjects with varying appearance, including facial hair and visible pigmented spots.

We additionally evaluate mesh depth MAE with and without the wrinkle-band loss.

SurfaceDepth MAE ↓ (mm)
Tracked mesh Vc\mathbf{V}_c2.26
Corrected mesh Vw\mathbf{V}_w (ours)1.69
−- wrinkle-band loss Lwb\mathcal{L}_{\text{wb}}2.10

Removing the wrinkle-band loss reduces geometric accuracy, supporting its contribution to geometric refinement. However, this result does not establish that all recovered high-frequency geometric structures correspond to true wrinkles.

AuthorsOpenReview ↗

Official Comment by Authors (2)

W4. Interpretation of the Ablation Results

LPIPS measures image similarity, while NCS in Table 4 measures agreement between rendered Gaussian normals and DAViD predictions. Removing Lalign\mathcal{L}_{\mathrm{align}} slightly improves LPIPS (0.075→0.0730.075\rightarrow0.073) but reduces NCS (0.912→0.8950.912\rightarrow0.895). These results reflect the role of alignment in preserving surface-normal consistency, without implying uniform gains in image quality.

Since this rendered-normal metric does not directly assess the corrected mesh, we complement Table 4 with the direct mesh evaluation in W1. Against the high-resolution MVS reference, the correction reduces Chamfer by approximately 11% and improves surface NCS without involving Gaussians. This comparison assesses mesh refinement separately from Gaussian alignment.


W5. View Consistency and Normal Visualization

We have prepared two additional camera-sweep videos: subject 104 performing EXP-8, and avatar 074 driven by the FREE sequence of subject 253. Both show RGB renderings and normal maps alongside the baselines under the same camera trajectory, allowing view consistency to be inspected as expressions evolve.

Regarding the irregularities around the open mouth and bangs, reconstructing detailed hair and mouth-interior geometry remains challenging due to thin structures and complex occlusions. Our method focuses on expression-dependent facial skin geometry and Gaussian alignment with the corrected surface. Our normal maps visualize Gaussian primitive orientations across the entire avatar, including hair and mouth-interior regions that are not explicitly reconstructed by the facial-skin corrective model and can exhibit irregular normal patterns.


Q2. Direct Surface Evaluation against a High-Resolution MVS Reference

In addition to the direct mesh comparison in response W1, we evaluate the held-out FREE sequence of subject 253 using frames sampled at fixed one-second intervals. We use the same evaluation setup as in W1.

We measure Chamfer distance and surface NCS over the FLAME face region. Surface NCS is computed directly against the MVS reference, rather than against DAViD predictions. For the final Gaussian avatars, we extract surfaces using the same TSDF extraction for both methods.

SurfaceCD ↓ (mm)Surface NCS ↑
Tracked FLAME mesh1.8420.910
MeGA facial mesh (depth-supervised)1.8110.912
Corrected mesh (ours, no depth supervision)1.7680.911
TexAvatars surface (TSDF)1.3390.905
Our surface (TSDF)1.2770.919

As shown in the table, our final avatar achieves lower CD and higher surface NCS than TexAvatars on the challenging held-out FREE sequence with unseen expressions.

We have uploaded a preliminary revised manuscript with additional experiments and clarifications highlighted in blue, and will continue to update it during the discussion period.

Reviewer_aBiwOpenReview ↗

Response to authors

Thanks for your additional explains and results, most of my concerns are addressed, I'll keep my original score.

We thank the reviewer for the thoughtful feedback and careful consideration of our additional results. Your suggestions have helped us strengthen the experimental validation and improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.