acceptodds
Under review as a conference paper at ICLR 2027

ChronoMesh: Two-Stage Age Editing for 3D Head Mesh

Abstract

Mesh-based head assets are widely used in production, but many recent aging methods target volumetric 3D representations that do not translate cleanly to UV material maps. We propose ChronoMesh, a two-stage framework for age editing of existing 3D head meshes that updates both global head shape and wrinkle-normal details while preserving identity. Because facial aging is inherently multi-scale, jointly optimizing geometry and texture under image-based supervision tends to favor local shading cues, under-expressing macro-scale deformation. Our two-stage design mitigates this issue by separating shape deformation from material editing. In the first stage, we predict an age-conditioned residual in a parametric head model to drive shape deformation toward a target age. During this stage, age-dependent proxy normal maps provide wrinkle cues consistent with the target age. In the second stage, we synthesize age-conditioned edits for UV-space intrinsic maps using a pre-trained StyleGAN with inversion-based latent editing, including wrinkle-normal synthesis. Both stages use image-based supervision through differentiable rendering. Experiments show that our edits preserve identity better than both image-space editors and mesh-reconstruction baselines while moving the rendered appearance toward the target age, with wrinkles represented explicitly in the normal maps rather than baked into diffuse reflectance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Age editing, 3D morphable model, head mesh

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LBXVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ChronoMesh, a two-stage practical approach to age editing while preserving facial identity and compatibility with conventional mesh-based rendering pipelines. The proposed approach separates facial aging into macro-scale geometric deformation and fine-scale material editing. The framework is trained and evaluated using FFHQ-UV-Intrinsics, with experiments demonstrating age-conditioned facial deformation and wrinkle synthesis across a wide range of target ages. Quantitative and qualitative comparisons suggest that the method achieves improved identity preservation and age control under its primary evaluation metrics while producing editable 3D mesh assets with explicit wrinkle-normal details.

Critical strengths

The paper addresses an practical problem in 3D character creation by enabling age editing directly on mesh-based head assets. Unlike image-based or neural-rendering-based aging methods, the proposed framework produces geometry and UV-space material maps that can be used in conventional rendering pipelines. This representation is particularly valuable for applications requiring explicit geometry, editable materials, and consistent relighting.

Also, the two-stage formulation provides a reasonable solution to the optimization interference between macro-scale shape deformation and high-frequency wrinkle synthesis.

Critical weaknesses

The proposed framework primarily combines existing techniques, including FLAME-based geometric deformation, StyleGAN2 latent editing, pivotal tuning inversion, differentiable rendering, and age-estimator-based supervision. In particular, Stage 1 employs a relatively straightforward residual prediction network in FLAME shape space, while Stage 2 largely extends the existing SAM framework to UV-space intrinsic maps using separate generators. Although the proposed two-stage training strategy and normal-map supervision are practically meaningful, the technical novelty of the individual components beyond their integration appears somewhat limited. The paper would benefit from a clearer distinction between its methodological contributions and existing approaches to geometric age transformation and latent-space face editing.

Although the proposed method achieves the lowest age error under MiVOLO, which is also used for training, its performance under DEX is weaker than that of several existing methods. Furthermore, the qualitative results do not consistently demonstrate clear improvements over the baselines, particularly considering that some of these methods operate in different domains and are not directly comparable to the proposed approach.

Questions for authors

Could the authors provide additional evaluation of a human perceptual study to demonstrate that the generated appearances accurately reflect their target ages?

Could the authors clarify what is meant by "consistent relighting" in Figure 1? The figure does not present results under different lighting conditions, making it difficult to assess whether the proposed method achieves consistent appearance across varying illumination settings.

3 comments
AuthorsOpenReview ↗

Response to Reviewer LBXV

Thank you for the review. The new evaluation (a second estimator, Qwen3-VL, and two additional 3D baselines) is summarized in the general response.

W1: novelty beyond the integration of existing techniques

We agree that the individual components are established, and the revision now says so. Sec. 5 separates what is reused from what is specific to our method.

  • Relative to geometric age transformation. Earlier methods learn aging trajectories of 3DMM coefficients, or mesh-to-mesh transformations, from 3D scans (Sec. 2.2). Stage 1 instead learns an identity-conditioned FLAME residual from image-based supervision only. This needs the age-dependent proxy normal maps; without them, the optimization distorts the geometry (Table 1 and Fig. 3, "w/o proxy normal").
  • Relative to latent-space face editing. SAM edits a face image. Stage 2 applies the invert-then-offset scheme to the diffuse, normal, and specular UV maps of the asset and adds a normal-map age estimator, which supervises wrinkles independently of the lighting (Table 1, "w/o normal loss": DEX 15.5 vs. 13.6, Qwen3-VL 19.8 vs. 15.9).
  • As a whole. The output is the input mesh asset with an edited shape and edited UV maps, not an image or a new neural representation.

W2: weaker results under DEX, and baselines from different domains

Age accuracy. We agree. MiVOLO is our training supervision, so its low error is not evidence of age accuracy. With DEX and with the newly added Qwen3-VL, neither of which is a training loss, our error is about 15 years, and several methods are more accurate in age. The revision states this in Sec. 4.2 and Sec. 6, and Table 2 no longer highlights our MiVOLO value. What our method retains is the highest identity similarity of all methods, also with ArcFace, which is not used in our training.

Different domains. To our knowledge, no existing method edits an existing mesh together with its diffuse, normal, and specular UV maps, so no baseline operates in exactly our domain. Sec. 4.2 now describes the baselines as three groups: (i) image-based editors applied to renders of the input mesh, which show what is achievable in image space but do not produce 3D assets; (ii) meshes reconstructed from the SAM and FADING outputs with ID-Sculpt; and (iii) two 3D editing methods, FaceG2E and GaussianAvatar-Editor, which are now part of the quantitative comparison. Among the 3D methods, GaussianAvatar-Editor and SAM+ID-Sculpt have lower age errors than ours but much lower identity similarity (ArcFace 0.467 and 0.209 vs. 0.786), and FaceG2E, whose identity similarity is closest to ours (0.741), has about twice our age error. FaceG2E and GaussianAvatar-Editor also require a separate optimization per subject and target age (about 1.5 and 7.7 hours per subject for 11 target ages, vs. about 1.6 minutes for ours).

Qualitative results. The comparison in Sec. 4.2 now reports the trade-off above instead of an overall "best" claim. Fig. 4 and Figs. 8 to 13 show the corresponding differences between the methods.

Q1: human perceptual study

We agree that a human perceptual study would provide valuable complementary evidence. Prior work has shown that human age judgments exhibit systematic biases and variability across observers and stimuli (Voelkle et al., 2012; Clifford et al., 2018). A study suited to our rendered faces therefore requires careful design and pilot testing to account for these effects. We were unable to complete such a study in time for this revision. Instead, we added an evaluation with Qwen3-VL, which was not used in training, to examine whether the age edits generalize beyond the training estimator. This additional automatic evaluation does not replace human perceptual validation, which remains a limitation of the present work.

Q2: "consistent relighting"

We meant that, because wrinkles are stored in the normal maps rather than baked into the diffuse map, the edited asset can be rendered under new lighting with a standard renderer and the shading of the wrinkles follows the light. We agree that the term was unclear. The revision replaces the term with "relightable" (Sec. 1, Sec. 4.2) and adds Fig. 17 with App. I, which show the edited assets of Fig. 1 rendered in Blender under four lighting conditions at target ages 10, 20, 60, and 80.

We would be glad to clarify any remaining points.

References

  • Voelkle, M. C., Ebner, N. C., Lindenberger, U., & Riediger, M. (2012). Let me guess how old you are: Effects of age, gender, and facial expression on perceptions of age. Psychology and Aging, 27(2), 265-277. https://doi.org/10.1037/a0025065
  • Clifford, C. W. G., Watson, T. L., & White, D. (2018). Two sources of bias explain errors in facial age estimation. Royal Society Open Science, 5(10), 180841. https://doi.org/10.1098/rsos.180841
AuthorsOpenReview ↗

Follow-up on Q1: human perceptual study

We have completed a human perceptual study, and the latest revised PDF includes it (Sec. 4.2, App. J). This replaces our answer to Q1 above, in which we wrote that we had not been able to complete such a study. The full summary is in our comment "Update: human perceptual study added to the revision"; in short:

  • Setup. Nineteen participants compared our method with the two 3D editing baselines, GaussianAvatar-Editor and FaceG2E, on 28 subjects and four target ages (10, 20, 60, 80). For each result, they estimated the age change relative to a render of the input mesh, and they ranked the three results by identity.
  • Identity. Our method is ranked first in 64% of the answers (GaussianAvatar-Editor 33%, FaceG2E 3%) and higher than both baselines (p < 0.0001). By target age, it is ranked first most often at 10 and 20 (86% and 99%), about as often as GaussianAvatar-Editor at 60 (48% vs. 51%), and less often at 80 (24% vs. 68%).
  • Age. The mean absolute errors are 17.2 years (ours), 18.4 (GaussianAvatar-Editor), and 27.1 (FaceG2E). Ours and GaussianAvatar-Editor do not differ significantly (p = 0.21), and both are lower than FaceG2E (p < 0.0001). Our results look 12.0 years younger than intended on average, which agrees with the DEX and Qwen3-VL results rather than with MiVOLO (cf. W2).
  • Caveats. The subjects were selected by the quality of the GaussianAvatar-Editor results, which may favor GaussianAvatar-Editor, and none of them is 50 or older.

While the overall direction is reasonable, the contribution and maturity of the manuscript are not yet sufficient for acceptance.

I believe the authors could further develop the methodology and provide more rigorous experiments to make the paper ready for a future venue.

Reviewer_Xu7NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ChronoMesh, a two-stage framework for 3D head mesh age editing. The first stage predicts age-conditioned FLAME shape deformation using an MLP, guided by proxy normal maps during training. The second stage fixes the geometry and edits UV diffuse, normal, and specular maps through StyleGAN-based latent editing. The method is evaluated through baseline comparisons, ablation studies, and mesh-space plausibility analysis.

Critical strengths
  1. The two-stage framework separates macro-scale shape deformation from UV material editing to mitigate optimization interference, where material changes can satisfy age supervision while macro-scale geometry deformation remains under-optimized. This provides a reasonable approach to addressing age-dependent changes at different scales.
  2. The method directly edits FLAME-based head meshes and their UV-space diffuse, normal, and specular maps, maintaining compatibility with standard graphics pipelines and supporting practical applications in 3D character editing.
Critical weaknesses
  1. Separating geometry deformation from material editing is an established approach, and the optimization interference between geometry and appearance is a known issue. Stage 1 uses an MLP to predict FLAME shape residuals, while Stage 2 largely adapts SAM-based latent editing to UV material maps. The overall methodological novelty is therefore limited.
  2. Although the authors provide qualitative ablations and mesh-space plausibility analysis, the experiments do not directly demonstrate that the two-stage approach produces more accurate age-dependent geometry than joint optimization. Furthermore, the substantial discrepancy between MiVOLO and DEX age-estimation results raises concerns about age-editing reliability, which could be better assessed through independent human perceptual evaluation.
Questions for authors

Q1. Could the authors clarify the main methodological differences between the proposed framework and existing two-stage geometry and appearance editing approaches?

Q2. How do the current ablation results support the claim that two-stage optimization mitigates under-optimization of geometry?

Q3. How do the authors explain the substantial discrepancy between MiVOLO and DEX age-estimation results in Table 2?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Xu7N

Thank you for the review. The new evaluation (a second estimator, Qwen3-VL, and two additional 3D baselines) is summarized in the general response.

W1 / Q1: novelty and differences from existing two-stage geometry and appearance editing

We agree that separating geometry from appearance, and the interference between the two, are known. Sec. 5 of the revision now states what is reused and what is specific to our method.

  • Reused: FLAME, SAM-style latent editing with e4e and PTI inversion, differentiable rendering, and age-estimator supervision.
  • Specific (1), the object that is edited. Methods that separate shape and appearance changes (Teng et al., 2024; Li et al., 2024) operate on reconstructed proxies or neural representations rather than on a mesh asset with UV maps. Among mesh-based methods, MeshWGAN edits geometry only, SemUV edits diffuse textures only, and FaceG2E does not edit normal or specular maps (Sec. 2.2). Our two stages edit an existing mesh asset itself: the shape as a FLAME residual, and the diffuse, normal, and specular UV maps as per-map latent offsets.
  • Specific (2), what makes the split trainable with image-based supervision only. In Stage 1, no age-dependent normal map exists yet, and the optimization compensates for age-inconsistent shading by distorting the geometry. Age-dependent proxy normal maps prevent this (Table 1, "w/o proxy normal": ArcFace 0.692 vs. 0.795; Fig. 3: distorted lips and chin). In Stage 2, a normal-map age estimator supervises wrinkles independently of the lighting (Table 1, "w/o normal loss": DEX 15.5 vs. 13.6, Qwen3-VL 19.8 vs. 15.9).

W2 / Q2: what the ablation shows about two-stage vs. joint optimization

What it shows. The jointly trained variant ("Single stage") satisfies the training estimator about as well as the two-stage model (MiVOLO 3.47 vs. 3.05), but its renders are farther from the target age for the two estimators that are not training losses (DEX 16.5 vs. 13.6, Qwen3-VL 21.3 vs. 15.9; Table 1). In Fig. 3, its geometry stays close to the source while its normal maps change with age, and its identity similarity is higher (ArcFace 0.858 vs. 0.798). These observations are consistent with the shortcut described in Sec. 1, in which joint training satisfies the age loss through the maps rather than through deformation.

What it does not show. There is no ground truth for the aged geometry of a given person, so we do not claim that the two-stage geometry is more accurate. The evidence on geometry is the qualitative comparison in Fig. 3 and the mesh-space analysis in Sec. 4.3: with respect to age- and sex-specific 3D statistics, the scores of our edited meshes approach those of unedited meshes from late adolescence onward. We revised Sec. 4.1 so that its statements stay within this scope.

W2 / Q3: discrepancy between MiVOLO and DEX, and human evaluation

MiVOLO is a training loss, so our networks fit it, and its low error (3.45) is not evidence of age accuracy. DEX and Qwen3-VL are not training losses and agree with each other (14.5 and 15.9 years), so we regard about 15 years as our age error. The revision states this (Sec. 4.2, Sec. 6), and Table 2 no longer highlights our MiVOLO value. For the other methods in Table 2, the three estimators differ by at most 6.3 years (e.g., GaussianAvatar-Editor: 10.2, 11.4, and 11.5), whereas for ours they differ by more than 12 years. SAM, whose age loss uses DEX, shows the same pattern with DEX (5.38 vs. 10.6 with MiVOLO and 7.84 with Qwen3-VL). The discrepancy is thus specific to the estimator used in training.

We agree that a human perceptual study would provide valuable complementary evidence. Prior work has shown that human age judgments exhibit systematic biases and variability across observers and stimuli (Voelkle et al., 2012; Clifford et al., 2018). A study suited to our rendered faces therefore requires careful design and pilot testing to account for these effects. We were unable to complete such a study in time for this revision. Instead, we added an evaluation with Qwen3-VL, which was not used in training, to examine whether the age edits generalize beyond the training estimator. This additional automatic evaluation does not replace human perceptual validation, which remains a limitation of the present work.

We would be glad to clarify any remaining points.

References

  • Voelkle, M. C., Ebner, N. C., Lindenberger, U., & Riediger, M. (2012). Let me guess how old you are: Effects of age, gender, and facial expression on perceptions of age. Psychology and Aging, 27(2), 265-277. https://doi.org/10.1037/a0025065
  • Clifford, C. W. G., Watson, T. L., & White, D. (2018). Two sources of bias explain errors in facial age estimation. Royal Society Open Science, 5(10), 180841. https://doi.org/10.1098/rsos.180841
AuthorsOpenReview ↗

Follow-up on W2 / Q3: human perceptual evaluation

We have completed the human perceptual evaluation suggested in Q3, and the latest revised PDF includes it (Sec. 4.2, App. J). This replaces the last paragraph of our response above, in which we wrote that we had not been able to complete such a study. The full summary is in our comment "Update: human perceptual study added to the revision"; the points related to your question are as follows.

  • Setup. Nineteen participants compared our method with GaussianAvatar-Editor and FaceG2E on 28 subjects and four target ages (10, 20, 60, 80). For each result, they estimated the age change relative to a render of the input mesh, and they ranked the three results by identity.
  • MiVOLO vs. DEX. The participants judged our results to be 12.0 years younger than intended on average, with a mean absolute error of 17.2 years. This is of the order of our DEX and Qwen3-VL errors (14.5 and 15.9), not of our MiVOLO error (3.45), and is consistent with our statement that the MiVOLO error is not evidence of age accuracy. The values are not directly comparable, because the subjects and target ages differ from those of Table 2.
  • Comparison with the 3D baselines. Our age error does not differ significantly from that of GaussianAvatar-Editor (18.4 years, p = 0.21) and is lower than that of FaceG2E (27.1 years, p < 0.0001). For identity, our method is ranked first in 64% of all answers (GaussianAvatar-Editor 33%, FaceG2E 3%) and higher than both baselines (p < 0.0001); at the target ages 60 and 80, however, GaussianAvatar-Editor is ranked first as often or more often (51% vs. 48% and 68% vs. 24%).
  • Caveats. The subjects were selected by the quality of the GaussianAvatar-Editor results, which may favor GaussianAvatar-Editor, and none of them is 50 or older.

I appreciate the authors' efforts in conducting the additional human perceptual study and evaluations, which address some of my concerns. However, I still share the concerns regarding limited methodological novelty and insufficient experimental validation. Therefore, I will maintain my current rating.

Reviewer_PFQROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

his paper presents ChronoMesh, a two-stage framework for age editing of 3D head meshes with UV-mapped intrinsic maps, designed to be compatible with standard production pipelines. The method explicitly decouples macro-scale shape deformation from wrinkle-level normal detail synthesis to avoid optimization bias toward local shading cues. In Stage 1, an MLP predicts an age-conditioned residual in FLAME shape space to drive geometric deformation, assisted by age-dependent proxy normal maps that provide consistent wrinkle shading cues. In Stage 2, UV-space intrinsic maps (diffuse, normal, specular) are edited via inversion-based latent offset in pre-trained StyleGAN2 models, with an auxiliary normal-map age estimator providing direct supervision for wrinkle learning. Both stages use differentiable rendering under a pre-optimized fixed lighting setup for image-based supervision.

Critical strengths

The problem is well-motivated for real production pipelines. Targeting mesh-and-UV assets directly, rather than neural 3D representations, makes the output directly usable in standard DCC tools and game engines, with clear practical value for digital character reuse.

Critical weaknesses
  1. Editing performance is underwhelming and raises severe overfitting concerns. On the unseen DEX age estimator, the method achieves a mean absolute error of 14.5 years, which is drastically worse than nearly all 2D baselines (e.g., SAM: 5.38, AgeTransGAN: 5.49). While MiVOLO error is low, this is the exact estimator used as training supervision. This strongly suggests the method overfits to biases of the MiVOLO model rather than producing genuine age transformation. Visually, extreme age edits (young children, advanced old age) appear unnatural, with anatomically imprecise deformation and flat, unconvincing wrinkle details.
  2. The training dataset has fundamental limitations in quality, diversity, and generality.
    • Ground-truth quality is insufficient. FFHQ-UV-Intrinsics provides reconstructed intrinsic maps from monocular images, not high-fidelity scanned UV textures. The diffuse, normal, and specular maps contain inherent reconstruction errors and artifacts, placing a low ceiling on the quality of supervision signals.
    • Age and demographic bias. The dataset is heavily adult-centric, with very few child and elderly samples, which directly limits extreme-age editing performance — a limitation the authors themselves acknowledge regarding FLAME’s shape space.
    • Topology lock-in. All data is aligned to a fixed FLAME topology, so the method cannot be applied to arbitrary artist-created meshes or raw scanned assets, severely restricting its practical scope.
    • Selection bias. Samples with failed FLAME fitting are discarded entirely, leaving only “easy” standard faces in the training set. Robustness on atypical face shapes is unvalidated.
  3. Fully decoupled two-stage training. The two stages are trained completely independently with no end-to-end joint refinement. Errors from Stage 1 propagate directly to Stage 2 without correction, leading to accumulated inaccuracy.
Questions for authors
  1. FFHQ-UV-Intrinsics provides reconstructed rather than scanned ground-truth intrinsic maps. How do you ensure the reliability of supervision signals for normal and diffuse maps? Have you validated results on high-resolution 3D scan datasets with artist-quality UV textures?
  2. DEX age error (14.5 years) is far higher than all 2D baselines, while MiVOLO error is low. Do you acknowledge that the method may be overfitted to the MiVOLO estimator used in training?
2 comments
AuthorsOpenReview ↗

Response to Reviewer PFQR

Thank you for the detailed review. The new evaluation (a second estimator, Qwen3-VL, and two additional 3D baselines) is summarized in the general response.

W1 / Q2: age accuracy and overfitting to MiVOLO

Yes, we acknowledge this. MiVOLO is our training supervision, and our low MiVOLO error (3.45) does not carry over to estimators that are not training losses: our error is 14.5 years with DEX and 15.9 with Qwen3-VL. The revision states this in Sec. 4.2 and Sec. 6, and Table 2 no longer highlights our MiVOLO value.

Whether the edits change the apparent age beyond what MiVOLO sees can be judged from the other two estimators:

  • With Qwen3-VL, the estimated age follows the target age (r = 0.861; GaussianAvatar-Editor 0.895, FaceG2E 0.527; Table 5).
  • Our age error is about half that of FaceG2E (DEX 14.5 vs. 30.5, Qwen3-VL 15.9 vs. 30.5), whose results change little with the target age.
  • In the ablation, the full model has lower DEX and Qwen3-VL errors than the single-stage variant and the variant without the normal-map loss (Table 1).

So the apparent age changes in the intended direction, but not far enough: with Qwen3-VL, our results are estimated 14.2 years younger than the target on average.

On the comparison with 2D editors: SAM's age loss uses the DEX classifier that we use for evaluation, so its DEX error (5.38) is not independent evidence either, and AgeTransGAN (5.49) is evaluated only on targets 0 to 70. With Qwen3-VL, most 2D editors are still more accurate in age than ours (e.g., SAM 7.84, Cradle2Cane 7.79), and the paper now says so. They output images rather than 3D assets, and their identity similarity is lower than ours (ArcFace 0.397 to 0.730 vs. 0.786); lifting the SAM and FADING outputs to meshes reduces it to 0.209 and 0.262.

Extreme ages. We agree that these are the weakest cases. Child targets are limited by the adult-centric shape space of FLAME, which the mesh-space analysis also shows (Sec. 4.3, Sec. 6), and the Qwen3-VL error is largest at targets 90 and 100 (Table 5, Sec. 6). Wrinkles are represented in the normal maps; the new Fig. 17 shows how their shading responds to four lighting conditions.

W2 / Q1: training data

  • Reliability of supervision. The maps of FFHQ-UV-Intrinsics are not regression targets of the age edits. They are used to train the StyleGAN2 generators of the UV maps without supervision and serve as input assets; the age supervision comes from rendered images and from the normal-map age estimator (Sec. 3.2 to 3.4). Their reconstruction errors therefore limit the fidelity that the generators can reach, which is the first limitation in Sec. 6.
  • Validation on scans. We have not evaluated on high-resolution scans with artist-quality textures. The closest check is the mesh-space analysis in Sec. 4.3, which compares the edited geometry with age- and sex-specific statistics of 3D facial growth curves; it concerns geometry only.
  • Age bias. We agree. The training data contains few subjects at the extremes of the age range, and FLAME's shape space is adult-centric (Sec. 6); both constrain extreme targets.
  • Topology. Correct: the input must follow the FLAME topology. Fitting FLAME to other meshes and transferring the deformation back is future work (Sec. 6).
  • Selection bias. Correct: subjects for which FLAME fitting failed are excluded from training and testing (about 5% of the 10,000 subjects; Sec. 3.5, App. A), so robustness on such faces is not evaluated.

W3: decoupled training

The "Single stage" variant in Table 1 trains shape and UV maps jointly. With both estimators that are not training losses, its age error is higher than that of the two-stage model (DEX 16.5 vs. 13.6, Qwen3-VL 21.3 vs. 15.9), which is why we separate the stages. We agree that Stage 2 cannot correct errors of Stage 1. Joint fine-tuning initialized from the two-stage solution is a natural extension, which we leave for future work.

We would be glad to clarify any remaining points.

AuthorsOpenReview ↗

Follow-up on W1 / Q2: human evaluation of age accuracy

The latest revised PDF adds a user study with 19 participants (Sec. 4.2, App. J; summarized in our comment "Update: human perceptual study added to the revision"). It bears on W1 / Q2 as follows.

  • Perceived age. The participants judged our results to be 12.0 years younger than intended on average, with a mean absolute error of 17.2 years. This is of the order of our DEX and Qwen3-VL errors (14.5 and 15.9), not of our MiVOLO error (3.45), and is consistent with our acknowledgment that the low MiVOLO error is not evidence of age accuracy. The values are not directly comparable, because the study covers 28 subjects and the target ages 10, 20, 60, and 80.
  • 3D baselines. Our age error does not differ significantly from that of GaussianAvatar-Editor (18.4 years, p = 0.21) and is lower than that of FaceG2E (27.1 years, p < 0.0001). For identity, our method is ranked first in 64% of all answers (GaussianAvatar-Editor 33%, FaceG2E 3%); at the target ages 60 and 80, however, GaussianAvatar-Editor is ranked first as often or more often (51% vs. 48% and 68% vs. 24%).
  • Scope. The study does not include the extreme target ages (0, 90, 100), and the source ages of its subjects are below 50, so it does not change our answer on extreme ages.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response: summary of the revision

We thank the reviewers for their constructive feedback and have uploaded a revised PDF. All three reviews question the gap between our MiVOLO and DEX age errors (PFQR W1/Q2, Xu7N W2/Q3, LBXV W2). We address this common point here and reply to the other points under each review. W and Q denote the Critical Weaknesses and Questions of each review, numbered in order of appearance.

Summary. With estimators that are not training losses, our age error is about 15 years, which is higher than that of several baselines, and we no longer present the MiVOLO result as evidence of age accuracy. Our method has the highest identity similarity of all compared methods, and the same estimators support the two-stage design over joint training.

1. Evaluation with estimators that are not training losses

MiVOLO supervises our training, so we agree that its low error is not evidence of age accuracy. The revision states this (Sec. 4.2) and no longer highlights our MiVOLO value in Table 2. We added the following.

  • Second estimator. Besides DEX, we now report Qwen3-VL (32B, Instruct), a vision-language model that is not used in our training. On real faces with age labels (UTKFace, CACD, AgeDB), its MAE is 4.6 to 7.1 years, within 0.7 years of MiVOLO on the three sets of face crops (App. H, Table 3).
  • 3D editing baselines. GaussianAvatar-Editor and FaceG2E are now part of the quantitative comparison.

3D rows of the revised Table 2 (age: MAE in years over target ages 0, 10, ..., 100):

MethodArcFace ↑BlendFace‡ ↑DEX ↓MiVOLO‡ ↓Qwen3-VL ↓
SAM+ID-Sculpt0.2090.10613.712.99.10
FADING+ID-Sculpt0.2620.15216.215.112.1
GaussianAvatar-Editor0.4670.35011.410.211.5
FaceG2E0.7410.60730.524.230.5
Ours0.7860.67214.53.4515.9

‡ Used in our training losses.

  • Age. DEX and Qwen3-VL agree: our error is about 15 years (14.5 and 15.9), not 3.45. Our method does not have the lowest age error: GaussianAvatar-Editor, SAM+ID-Sculpt, and most 2D editors are lower on both estimators, whereas FaceG2E has about twice our error.
  • Identity. Our identity similarity is the highest of all ten methods, also with ArcFace, which is not used in our training. The 3D methods with lower age errors have much lower identity similarity (ArcFace 0.467 and 0.209 vs. 0.786).
  • Trend and bias. With Qwen3-VL, the estimated age follows the target age (Pearson r = 0.861; GaussianAvatar-Editor 0.895, FaceG2E 0.527), but our results are estimated 14.2 years younger than the target on average, and the error is largest at targets 90 and 100 (App. H, Table 5). Sec. 6 now lists the gap between MiVOLO and the other estimators as a limitation.
  • Sensitivity. Qwen3-VL results depend on the model variant and prompt. Over the settings in App. H (Table 4), excluding one in which many answers were invalid, our MAE ranges from 15.9 (the setting of the main paper, chosen on the labeled datasets) to 19.9. In all eight settings, GaussianAvatar-Editor has the lowest error of the three 3D editing methods, followed by ours and FaceG2E.

2. Two-stage vs. single-stage training (PFQR W3, Xu7N W2/Q2)

Table 1 now includes Qwen3-VL:

VariantArcFace ↑DEX ↓MiVOLO‡ ↓Qwen3-VL ↓
Single stage (joint)0.85816.53.4721.3
Two-stage (ours)0.79813.63.0515.9

With the two estimators that are not training losses, the two-stage model has the lower age error, while MiVOLO barely separates the two. The single-stage model has higher identity similarity.

3. Other changes

  • Sec. 4.2: the baselines are described as three groups (image-based editors, meshes reconstructed from their outputs, and 3D editing methods), and the comparison reports the trade-off between age error and identity instead of a "best" claim.
  • Sec. 5: which components are reused and which are specific to our method (Xu7N W1/Q1, LBXV W1).
  • "Consistent relighting" is replaced by "relightable"; Fig. 17 and App. I show the edited assets under four lighting conditions (LBXV Q2).
  • Sec. 4.1: statements on the single-stage variant are limited to what Table 1 and Fig. 3 show.
  • App. H: Qwen3-VL prompts, decoding settings, validation, sensitivity, and errors by target age.
AuthorsOpenReview ↗

Update: human perceptual study added to the revision

This comment replaces the statement in our responses to Reviewers Xu7N (W2/Q3) and LBXV (Q1) that we had not been able to complete a human perceptual study. We have now completed one, and the latest revised PDF includes it (a paragraph in Sec. 4.2, and App. J with Tables 6 and 7). Below is a summary, including the results that do not favor our method. Compared with the PDF that accompanied our earlier responses, the other changes are limited to the Ethics statement, which now covers the study, and minor wording in Sec. 4.3.

Summary. For identity, our method is ranked highest on average (first in 64% of the answers), owing to the target ages 10 and 20; at 60 it is on par with GaussianAvatar-Editor (GA-Editor below), and at 80 GA-Editor is ranked first more often. For age, ours and GA-Editor do not differ significantly (17.2 vs. 18.4 years), both have lower errors than FaceG2E (27.1), and the results of all three methods look younger than intended.

Design

  • Participants. 19 adults who were not involved in this research, with consent and without compensation; the Ethics statement gives the details and the assessment of the need for an ethics review.
  • Questions. Ours is compared with the two 3D editing baselines of Table 2, GA-Editor and FaceG2E, on 28 test meshes and four target ages (10, 20, 60, 80). Each question shows a render of the unedited input mesh as the reference and the three results for one target age, labeled A, B, and C in random assignment; the target age is not shown. Four forms assign the target ages to the meshes in a Latin square, and each participant answered one form of 28 questions (532 answers in total).
  • Answers. (1) For each result, how many years older or younger the person looks than in the reference (−80 to +80 in 5-year steps); the age error is this answer minus the intended change (target age minus source age). (2) A ranking of the three results by how much each looks like the same person as the reference.
  • Input mesh selection. GA-Editor gives blurry or color-shifted results for some input meshes. The 28 input meshes were drawn at random from the 39 test meshes whose GA-Editor results passed a sharpness and color check at all eleven target ages, which may favor GA-Editor.

Results

MeasureOursGA-EditorFaceG2E
Identity: mean rank (1 = best)1.411.862.74
Identity: ranked first (% of answers)64.332.53.2
Age: mean absolute error (years)17.218.427.1
Age: mean signed error (years)−12.0−15.8−20.8
Age: smallest absolute error (% of answers)33.143.723.2

Ranked first for identity by target age (% of answers; Table 7):

Target ageOursGA-EditorFaceG2E
1085.711.33.0
2099.20.00.8
6048.151.10.8
8024.167.78.3
  • Identity. Over all answers, our method is ranked higher than both baselines (mixed-effects model, Holm-adjusted p < 0.0001; Wilcoxon signed-rank tests per participant and per mesh agree; Table 6).
  • Age. Our absolute error is lower than that of GA-Editor by an estimated 1.27 years (95% CI [−0.70, +3.24]), which is not significant (p = 0.21). Both have lower errors than FaceG2E (p < 0.0001). By target age, the estimates favor ours at 20 (by 8.1 years) and GA-Editor at 60 (by 5.1 years).

Relation to the automatic metrics

  • Age. The participants judged our results to be 12.0 years younger than intended on average, the same direction as with Qwen3-VL (14.2 years younger than the target; Table 5). Our mean absolute error of 17.2 years is of the order of our DEX and Qwen3-VL errors (14.5 and 15.9), not of our MiVOLO error (3.45). The values are not directly comparable, because the study covers 28 meshes and four target ages and measures the age change relative to the reference.
  • Identity. As with ArcFace and BlendFace, our method is rated highest on average. The order of the two baselines differs: the participants rank GA-Editor above FaceG2E, whereas both networks give FaceG2E the higher similarity (Table 2). The study covers only four target ages and only meshes whose GA-Editor results passed the sharpness and color check, whereas Table 2 is not restricted in these ways; this may contribute to the difference, but we have not verified it (App. J).

Limitations. The study has 19 participants, 28 meshes, four target ages, and two baselines. Of the 28 meshes, 20 are between 20 and 39 and none is 50 or older (the source ages are the dataset's age labels, estimated from the photographs), so the study mainly evaluates edits of faces in their twenties and thirties.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.