acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Task-Defined Representation Alignment and Calibrated Exchange for Controllable Fingerprint Generation

Abstract

Controllable fingerprint generation offers a promising way to expand limited biometric datasets and support presentation-attack research. Beyond visual realism, its practical utility depends on modifying a specified factor while preserving identity-related ridge patterns, non-target attributes, and fine-grained image details. This is challenging because fingerprint content, presentation material, and acquisition batch are often correlated and expressed through shared, spatially distributed visual patterns. We propose TRACE (Task-Defined Representation Alignment and Calibrated Exchange), a two-stage framework that separates factor-structured representation learning from generation-oriented refinement. In the first stage, Factor-Structured Information Extraction combines factor-wise supervised contrastive learning with image reconstruction to learn semantically aligned latent groups that retain the information needed for faithful generation. In the second stage, Structured Latent Distillation freezes the learned encoder and calibrates these groups through residual refinement and factor-specific spatial adaptation, enabling selective factor exchange while limiting unintended changes to non-target attributes. Experiments on a real-world fingerprint dataset and Shapes3D show that TRACE achieves state-of-the-art factor organization and controllable generation while maintaining competitive reconstruction quality. These results establish TRACE as an effective solution for task-defined controllable generation, combining accurate and interpretable factor control with faithful image synthesis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
disentanglement representation learning, controllable image generation, contrastive representation learning

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_TAxnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper focuses on fingerprint generation that is controllable by selected task-defined factors. The challenge is to modify the target attributes without altering the ridge strcutures and other correlated semantics of the fingerprint. A two-stage approach is introduced in the paper: the first stage learns latent representations while the second stage disentangles factors and exchanges selected factors through residual refinement. Experiments on two datasets show the effectivenss of the proposed method.

Critical strengths
  1. The motivation of the control problem is explained clearly. The paper clearly distinguishes selective attribute control from visual quality.

  2. The paper is well-structured and easy to follow.

  3. The authors show that their proposed method can also generalize to Shape3D.

Critical weaknesses
  1. There are similar attribute disentanglement and synthesis methods from other attribute control tasks [1][2] except they use GAN. This paper mainly focuses on the fingerprint generation task, but I did not see a model design to address challenges specific for the fingerprint task.

  2. The metrics lack an explicit evaluation on minutiae preservation and fingerprint identity preservation. The reported metrics are reconstruction scores, DCI, and classifier-based CFSS, which do not directly establish preservation of minutiae or fingerprint matchability. I suggest the authors to refer to GenPrint paper on the evaluation.

  3. The authors mentioned that controllable fingerprint generation can be used to synthesize data in some downstream tasks where real data are difficult to acquire. I suggest the authors add an experiment on one such downstream task using the fingerprint images synthesized by their method to evaluate its generalization ability.

Questions for authors

In the introduction, the authors address the importance of keeping the ridge strcuture unchanged after controllable generation. Is there a specific design choice in the model that can improve the preservation of the ridge structure? If there is, can the authors provide ablationon that?

1 comment
AuthorsOpenReview ↗

Clarification of the motivation of model design, identity-related structure preservation, and downstream utility

We sincerely thank you for your time and effort in reviewing our manuscript and for your insightful comments. We appreciate your positive assessment of the clear motivation, the distinction between selective attribute control and visual quality, the paper’s organization, and the applicability of our approach to Shapes3D. Your suggestions have helped us clarify our design rationale and strengthen the evaluation of identity-related structure preservation and downstream utility. We address your comments below:

1. Clarification of the motivation behind the model design

Thank you for your insightful comments. We address the connection between our model design and the fingerprint task, as well as the role of this design in ridge structure preservation below.

(1) Motivation of model design . Our design is motivated by the fingerprint control task: modifying a target attribute, such as material, while preserving identity-related ridge structure and other non-target information. To support this goal, TRACE learns factor-aligned latent blocks from task-defined labels and generates images by explicitly swapping the target block. Thus, the connection to the fingerprint task lies in the factor-structured representation and selective exchange mechanism, rather than a fingerprint-specific architecture. The results on Shapes3D further demonstrate that this mechanism is also applicable beyond fingerprints.

We could not locate the bibliographic details for [1] and [2] in the review and would appreciate clarification to enable a precise comparison.

(2) Clarification of design for ridge structure preservation and ablation. We do not introduce a ridge-specific choice in our model. In the fingerprint setting, the content block is intended to capture identity-related ridge structure, while exchanging the material block is intended to modify material-related characteristics while retaining the source image’s identity-related structure. Accordingly, there is no ridge-specific component to isolate in a dedicated ablation. To assess preservation empirically, we have added a SIFT-based keypoint comparison protocol in the revised manuscript in Appendix A.6. These results provide complementary evidence of identity-related local structure preservation.

2. Clarification of identity-related structure preservation

Thank you for your valuable comment. Commonly used fingerprint identity preservation methods, such as NFIQ and minutiae detection in GenPrint, are typically evaluated on larger fingerprint images (e.g., 388 × 374, 296 × 560, or 300 × 300 pixels in FVC2002 database). In contrast, our dataset consists of narrow, 160 × 26 partial fingerprints. Directly applying existing recognition methods to these images yields poor recognition performance for all compared baselines.

We therefore applied a SIFT-based keypoint comparison protocol to assess the preservation of identity-related local structures after swapping. For material swaps, our model achieves the second-highest score (0.463) for material-swapped images and the highest score (0.559) for batch-swapped images . The detailed algorithm and results are updated in Tables 2 and A9, and Appendix A.6 in the revised version.

3. Clarification of downstream utility and generalization

Thank you for your constructive suggestion. We have added a downstream material classification experiment in Section 4.3 and Appendix A.7 in the revised version, using images generated through TRACE’s material swapping to augment the training set. The results show that standard data augmentation, including flipping, rotation, translation, brightness/contrast jitter, and random erasing, improves classification performance on held-out real fingerprint images, while TRACE-based augmentation yields greater improvements. Specifically, TRACE achieves 75.51% accuracy, 74.12% balanced accuracy, and 72.44% macro F1, exceeding standard augmentation by 2.23, 1.78, and 2.28 percentage points, respectively. These findings support the generalization ability of TRACE-generated data in this downstream task and suggest its potential to reduce the costs of sample fabrication, image collection, and dataset maintenance.

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

The paper proposes TRACE, a two-stage framework for controllable fingerprint generation. The first stage learns factor-aligned latent representations using contrastive learning and reconstruction, while the second stage refines these representations for selective factor swapping. Experiments on fingerprint and Shapes3D datasets show improvements in disentanglement and factor control compared to existing methods.

Critical strengths
  1. Controlling specific fingerprint attributes while preserving other important information is an interesting problem.
  2. The authors compare their method with several baselines and provide ablation studies to evaluate the contributions of different components.
Critical weaknesses
  1. Both stages seem to have similar goals, as they both try to learn factor-specific representations while maintaining reconstruction quality. The motivation for using two separate stages is not very clear. In Table 2, adding Stage II improves CFSS only slightly, from 0.3378 to 0.3580, while SSIM drops from 0.8288 to 0.607.
  2. The proposed CFSS metric mainly relies on separately trained, frozen classifiers to evaluate factor swapping. However, a high classification-based score does not necessarily mean that important fingerprint details, particularly identity-related ridge structures, are preserved.
  3. Section 4.4 states that removing contrastive supervision degrades both controllability and reconstruction fidelity. However, Table 2 shows that reconstruction actually improves substantially without InfoNCE, with SSIM increasing from 0.607 to 0.914 and MSE decreasing from 0.0127 to 0.0016. This contradicts the authors' interpretation of the results.
Questions for authors
  1. How well does the proposed method preserve actual fingerprint identity after factor swapping? Could the authors evaluate this using a fingerprint-matching system?
3 comments
AuthorsOpenReview ↗

Clarification of motivation of model design, identity-related structure preservation

We sincerely thank you for your time and effort in reviewing our manuscript and for your insightful comments. We appreciate your recognition of the importance of controlling specific fingerprint attributes while preserving other relevant information, as well as your positive assessment of our baseline comparisons and component-wise ablation studies. Your feedback has helped us clarify the motivation for the two-stage design, improve the interpretation of the ablation results, and strengthen the evaluation of identity-related structure preservation. We address your comments below:

1.Clarification of the motivation of two-stage design

Thank you for your thoughtful comments. Although both stages involve factor-specific representations and reconstruction, they serve different roles. Stage I uses reconstruction and contrastive supervision to learn an information-preserving, factor-aligned latent representation. Stage II then applies residual refinement to address the correlated information that remains in this representation and improve selective factor exchange. The motivation for the two-stage design is therefore to separate initial representation learning from subsequent refinement for factor swapping, but not to repeat the same learning objective.

Adding Stage II increases CFSS from 0.3378 to 0.3580 with decrease in SSIM from 0.8288 to 0.607. We interpret this result as a trade-off between controllability and reconstruction fidelity. As our primary objective is to improve controllable generation, therefore, we report the operating point with the highest CFSS on the Pareto frontier shown in Figures 6 and A6. At this operating point, TRACE still maintains competitive reconstruction quality relative to the evaluated baselines, as shown by the comparisons in Table 1 and the visualizations in Figure A2.

2. Clarification of identity-related structure preservation and fingerprint matching evaluation

Thank you for your valuable comment. Commonly used fingerprint identity preservation methods, such as NFIQ and minutiae detection, are typically evaluated on larger fingerprint images (e.g., 388 × 374, 296 × 560, or 300 × 300 pixels in FVC2002 database). In contrast, our dataset consists of narrow, 160 × 26 partial fingerprints. Directly applying existing recognition methods to these images yields poor recognition performance for all compared baselines.

We therefore applied a SIFT-based keypoint comparison protocol to assess the preservation of identity-related local structures after swapping. For material swaps, our model achieves the second-highest score (0.463) for material-swapped images and the highest score (0.559) for batch-swapped images . The detailed algorithm and results are updated in Tables 2 and A9, and Appendix A.6 in the revised version.

3. Correction of the interpretation of the InfoNCE ablation results

Thank you for your careful reading and for pointing out this inconsistency. Our original wording did not clearly convey the intended distinction. Removing InfoNCE improves reconstruction (SSIM: 0.607 → 0.914; MSE: 0.0127 → 0.0016) but substantially reduces CFSS (0.3580 → 0.0957). These results indicate that contrastive supervision supports controllable factor exchange at the cost of reconstruction fidelity in this ablation. We have corrected Section 4.5 in the revised version to accurately reflect these findings. Thank you again for your valuable comment!

Reviewer_YdNZOpenReview ↗

Response to rebuttal

  1. According to Table 4, FSIE only achieves a CFSS of 0.3378 with an SSIM of 0.829. This already exceeds every baseline in Table 1 on CFSS. Adding Stage II raises CFSS by about 0.02, which is close to the reported standard deviations, while SSIM drops to 0.607. Given these numbers, the empirical case for Stage II remains weak.
  2. Thank you for adding the SIFT analysis. However, it’s not clear that TRACE preserves identity better than the baselines: in Table 2, XFactors scores higher than TRACE on material swap. Also, SIFT similarity only compares local texture, so it doesn’t directly show that the fingerprint identity is preserved.

I appreciate the authors’ efforts, but my main concerns remain, so I will keep my current score.

AuthorsOpenReview ↗

Clarification of the contribution of Stage II and the evidence for preserving identity-related structures

Thank you for the follow-up and for considering our additional evaluation. We would like to clarify the contribution of Stage II and the evidence for preserving identity-related structures.

1. Stage II calibrates representations for selective factor exchange.

The strong performance of FSIE is an important result of our framework: Stage I already learns a useful, factor-aligned representation. However, aligning a latent block with its designated factor does not ensure that exchanging this block changes only that factor. Stage II addresses this distinction through factor preservation, conditional suppression of non-target information, and latent anchoring, while keeping the Stage-I encoder frozen.

The supporting evidence extends beyond the FSIE-to-TRACE comparison. In Table 4, removing conditional suppression and laten anchoring reduces CFSS from 0.3580 to 0.3103, removing all three refinement losses reduces it to 0.2838, and joint optimization yields only 0.1540. These ablations support explicit refinement and the separation of representation learning from exchange calibration, but not simply adding another reconstruction stage.

The SSIM decrease reflects the trade-off between reconstructing an unchanged input and selectively modifying a factor. Figures 6 and A6 characterize this trade-off; our reported configuration prioritizes CFSS for controllable generation. Stage II’s contribution is therefore improved selective exchange, but not uniformly better reconstruction.

2. The SIFT analysis evaluates local structures appropriate to partial fingerprints, with complementary visual evidence.

As noted in our previous response, our images are narrow 160 × 26 partial fingerprints, but not full fingerprint impressions. Direct application of conventional fingerprint-matching methods performs poorly across the evaluated methods in this setting. We therefore use SIFT to assess the identity-related local ridge structures actually visible in these partial observations. Specifically, our protocol compares descriptors at corresponding source-image locations and orientations, measuring how well local ridge appearance is retained after swapping.

The quantitative results show that TRACE is second-best for material swapping, behind XFactors (0.463 vs. 0.579), and best for batch swapping (0.559, compared with 0.457 for XFactors). These results demonstrate consistently strong local-structure preservation across both swap types.

Importantly, the SIFT scores are complemented by Figure A5, where TRACE produces clearer, more sharply defined ridge patterns than the evaluated baselines, including XFactors, in the displayed examples. The quantitative and qualitative evaluations provide complementary evidence: SIFT measures local descriptor similarity, while Figure A5 illustrates the clarity and definition of the generated ridge patterns.

Our intended conclusion is therefore that TRACE combines strong factor-swapping performance with competitive preservation of identity-related local structures, not that SIFT alone establishes biometric identity. We hope this clarifies both the task-specific value of Stage II and the evidence provided by the additional analysis. We thank you again for your thoughtful comment.

Reviewer_d8UfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a two-stage VAE for controllable fingerprint generation. Stage I learns factor-aligned latent blocks (content, material, batch) using per-block supervised contrastive loss and full-latent reconstruction. Stage II freezes the encoder and refines the blocks with factor classifiers for selective factor swapping. Evaluation uses a private spoof fingerprint dataset and a grouped Shapes3D subset.

Critical strengths

The problem of selective factor control for fingerprint data with correlated acquisition factors is relevant for biometrics.

Critical weaknesses

The main weakness is that the biometric utility is discussed but it is never measured. There are several other issues and in the current form, the paper is not ready for publication. Detailed feedback is in the next section.

Questions for authors
  • For a fingerprint paper I expected a matcher-based identity test after material swaps, and a PAD augmentation experiment on held-out real data with bona fide samples and unseen materials.
  • From biometrics point of view, what sensor and resolution produced these images, and how were the spoofs fabricated? Further, do impressions from the same finger or the same physical spoof appear across splits?
  • The paper does not evaluate the correlated-factor. Details are missing: which content, material and batch combinations exist in the data? The paper should evaluate swaps that land on rare or held-out combinations.
  • Each residual refiner sees the whole latent vector. Once one block is swapped, what stops the other blocks from restoring the original factor?
  • In Stage II, CFSS is raised from 0.338 to 0.358 and drops SSIM from 0.829 to 0.607. With reconstruction loss alone it is worse than FSIE-only on both. What is Stage II actually buying?
  • “Competitive reconstruction” is a unclear. Table 1 gives fingerprint SSIM of 0.597 against 0.913 for DyGA and 0.827 for XFactors. Further, ridge detail should be checked directly.
  • Several baselines are unclear and it looks like that they are untrained results. On Shapes3D, InfoDiffusion and FDAE reach SSIM 0.999 with CFSS near zero, which suggests the swapped block never reaches the decoder. Please check convergence, and explain how blocks are assigned to factors for unsupervised methods without touching test labels.
  • Stage I is close to XFactors, and Stage II descends from adversarial disentanglement (Mathieu et al., Hadad et al.). Fingerprint synthesis work (PrintsGAN, SpoofGAN, GenPrint) is also missing. Please position TRACE against these and compare where it makes sense.
  • Does DCI use Stage-I means or Stage-II refined blocks? It is identical across every SLD ablation, so Stage II’s effect on the representation is invisible.
  • Appendix A.1 computes DCI and the swap evaluation on the full 100,000-sample Shapes3D population, training samples included. Please report test-only numbers.
  • With βKL = 10⁻⁶ this is nearly a deterministic autoencoder. The paper should report posterior variances and compare against a plain autoencoder to show the stochastic encoding matters.
  • Were hyperparameters chosen on validation CFSS? Table 1 is seed 0 and Table 2 is a six-seed mean (SSIM 0.597 versus 0.607), so run the key comparisons across seeds and say which is which.
  • Several references need to be fixed. Here are a few examples: (i) GenPrint is TPAMI 47(2), 2025 (ii) Chugh and Jain, TIFS vol. 16, is 2021.
  • Minor: header of the paper says ICLR 2026 - not 2027.
4 comments
AuthorsOpenReview ↗

Clarification of biometric utility, identity-related structure preservation, and PAD evaluation

We sincerely thank you for your time and effort in reviewing our manuscript and for your detailed and insightful feedback. We particularly appreciate your recognition that selective factor control for fingerprint data with correlated acquisition factors is relevant to biometrics. Your valuable comments have helped us improve the clarity of our manuscript, we address your comments by topic.

Q1 and Q6: Biometric utility, identity-related structure preservation, and PAD augmentation

Thank you for your insightful comments. We have added a quantitative identity-preservation evaluation to the revised manuscript and clarify the following points.

  • Biometric utility evaluation. Conventional fingerprint quality assessment (e.g., NFIQ) and minutiae-based matching are typically applied to larger images (e.g., 388 × 374, 296 × 560, or 300 × 300 pixels in FVC2002 database), making them unsuitable for our small partial fingerprints (160 × 26 pixels). We therefore use SIFT-based keypoint matching to assess the preservation of identity-related local structures after swapping. The evaluation protocol is detailed in Appendix A.6 of the revised manuscript.

  • Matcher-based identity test. We evaluate identity preservation using SIFT-based keypoint matching. Our model achieves the second-highest score (0.463) for material-swapped images and the highest score (0.559) for batch-swapped images . More results are reported in Tables 2 and A9 of the revised version.

  • PAD augmentation experiment. Our study focuses on controllable fingerprint generation, thus our dataset does not include bona fide samples. Therefore, the PAD augmentation experiment is not available in our case.

AuthorsOpenReview ↗

Clarification of data acquisition, factor combinations, data splits, and evaluation scope

Q2: Fingerprint acquisition, spoof fabrication, and data splits

Thank you for your insightful comments.

  • Fingerprint acquisition. To preserve anonymity during double-blind review, we respectfully defer disclosure of the detailed acquisition setup, including the sensor model and native sampling resolution (PPI), as these details could identify the data collection source. We will disclose this information in the final version.

  • Spoof fabrication. Different suppliers fabricated physical fingerprint replicas using resin, gelatin, soft silicone, hard silicone, and wood glue. All molds were taken from genuine human fingers following the same molding procedure, and all replicas were imaged using the same sensor and acquisition protocol.

  • Data splits and evaluation scope. As described in Appendix A.1, we randomly partitioned the dataset at the image level using seed 42, without enforcing separation by finger identity or physical spoof instance. Our evaluation focuses on disentanglement and selective factor exchange within the dataset’s annotated factor categories, rather than generalization to unseen finger identities or physical spoof artifacts. This scope is consistent with the in-distribution evaluation setting commonly adopted in disentanglement research, such as Ada-ML-VAE, EncDiff, DisDiff, and XFactors.

Q3: Clarification of factor combinations and correlated-factor evaluation

Thank you for this helpful suggestion. As described in Appendix A.1, different batches contain different content identities, creating a correlated content–batch setting. Sampling is approximately balanced within each batch, with similar sample counts across identities, rather than being designed around rare combinations.

To further assess controllability in this setting, we have added a stratified swap evaluation in the revised Appendix A.5, grouping source–reference pairs by the number of differing factors (k = 1, 2, or 3). This analysis tests increasingly challenging exchanges involving content, material, and batch. Although controllability decreases as more factors differ, TRACE achieves the best overall performance among all evaluated baselines, demonstrating its effectiveness in factor exchange within this correlated setting.

Q10: Clarification of the Shapes3D data split and DCI evaluation

Thank you for your comment. The 100,000 is the sample size we used in Shapes3D experiments. As described in Appendix A.1, we split this subset into 64,000 training, 16,000 validation, and 20,000 test images using a split seed of 2025. DCI dimensions are constructed using the training set and evaluated on the test set. To avoid ambiguity, we have clarified this description in the revised Appendix A.1.

AuthorsOpenReview ↗

Clarification of model design, stage-II benefits, and relation to prior work

Q4: Clarification of cross-block interference

Thank for your comments. The SLD objectives are designed to stop this interference. During training stage. Conditional suppression (Eq. 10) directly stops other blocks from restoring the original block: it discourages other blocks from encoding information beyond their own labels. Factor preservation (Eq. 9) also maintains each block’s designated factor. These objectives constrain what each block retains. Swap generation only happens in the inference stage; therefore, no such interference happens.

Q5 and Q6: Clarification of stage-II benefits

Thank you for your insightful comments. Our primary objective is to improve controllable generation, so we select the operating point with the highest CFSS on the Pareto frontier shown in Figures 6 and A6. Stage II refines the correlated information present in the latent representations learned in Stage I, thereby improving CFSS. Despite a modest decrease in reconstruction performance, TRACE remains competitive with all baseline methods, as shown by the comparisons in Table 1 and the visualizations in Figure A2. Although its reconstruction scores are slightly lower than those of DyGA and XFactors, the three models produce nearly identical visual quality.

Q8: Clarification of TRACE’s positioning relative to prior work

Thank you for this comment. In Stage I, we use a reconstructive bottleneck: factor-aligned blocks jointly retain the visual information needed for faithful reconstruction. In contrast, XFactors adopts a disentangled information bottleneck with dedicated factor subspaces and a separate residual subspace to induce factor-exclusive compression. In Stage II, we do not adopt adversarial disentanglement; rather, we use adversarial training specifically to suppress non-target information beyond what is predictable from each block’s own factor label. This design targets factor-swap generation.

Our work also differs from PrintsGAN, SpoofGAN, and GenPrint in its research objective. These methods synthesize fingerprint images through rendering or conditioning, whereas TRACE generates images by explicitly exchanging latent factor blocks. Given this difference in objectives and generation mechanisms, we do not include them as baselines.

Q11: Clarification of the role of stochastic encoding

Thank you for your comment. TRACE targets factor-controllable fingerprint generation through factor exchange, but not generation through stochastic sampling. Posterior sampling is used during reconstruction training, while factor exchange operates on posterior-mean blocks.

AuthorsOpenReview ↗

Clarification of baseline implementation, DCI evaluation, experimental settings, and the format

Q7: Clarification of baseline training and evaluation

Thank you for your insightful suggestion. We have rechecked and confirmed that both InfoDiffusion and FDAE converge and that the swapped representations reach their decoders. In Table 1, we used t=50t=50 sampling steps, following DisDiff, for a fair comparison. Using InfoDiffusion’s T=100T=100 setting increases its CFSS from approximately zero to 0.004. For FDAE, we applied the same CFSS evaluation procedure used for all baselines to ensure a fair comparison. Restoring FDAE’s prescribed semantic block extraction step increases its CFSS from approximately zero to 0.466. The revised manuscript reports these updated results.

Unsupervised baselines learn block semantics through their original objectives, without factor-label supervision. For example, EncDiff maps latent blocks to concept tokens and learns their semantics through cross-attention-conditioned diffusion reconstruction. For swap evaluation, block–factor correspondences are determined using the validation set and fixed before testing. Test labels are never used to select or assign blocks.

Q9: Clarification of DCI evaluation in the SLD ablation study

Stage I is designed for factor disentanglement, whereas Stage II targets controllable generation. Accordingly, in the SLD ablation study, we compute DCI using the Stage-I means to evaluate disentanglement performance.

Q12: Clarification of hyperparameter selection and seed settings

Thank you for careful comments. Hyperparameters were selected based on validation-set CFSS in the sensitivity analyses shown in Figures 6 and A6.

In Table 1, all baselines were reproduced using seed 0, so we also report TRACE’s seed-0 results for a fair comparison. In Table 4, we report ablation results averaged over six training seeds to assess the robustness of each component.

Q13: Clarification of reference details

Thank you for your comments. All reference entries were consistently imported from Google Scholar’s BibTeX records.

Q14: Correction of the conference header

Thank you for your comment. We have updated the header to ICLR 2027 in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response summary to all reviewers and area chair.

Dear Reviewers and Area Chair,

We sincerely thank you for your time and constructive feedback. We appreciate your recognition of our motivation, empirical comparisons, ablation studies, and applicability beyond fingerprints. TRACE’s central contribution is to separate factor-aligned representation learning from calibration for selective factor exchange, addressing the distinction between identifying an attribute in a representation and changing it without disturbing non-target content.

Our revisions, highlighted in blue in the revised version, strengthen three aspects of the paper:

  • Design rationale and supporting ablations (d8Uf, YdNZ, TAxn). We clarify how Stage II complements Stage I through factor preservation, conditional suppression, and latent anchoring. The ablations support these mechanisms: removing conditional suppression and laten anchoring reduces CFSS from 0.3580 to 0.3103, while joint optimization yields 0.1540. We also clarify the controllability–reconstruction trade-off underlying our selected configuration (Section 4.4; Table 4; Figures 6 and A6).

  • New evaluation of local ridge-structure preservation after swapping (d8Uf, YdNZ, TAxn). For our narrow 160 × 26 partial fingerprints, we add a SIFT-based evaluation of identity-related local structures. TRACE ranks second on material swaps (0.463) and first on batch swaps (0.559). Complementary visual comparisons show clearer, more sharply defined ridge patterns than the evaluated baselines, including XFactors. We distinguish this evidence from biometric identity verification (Table 2; Figure A5; Appendix A.6).

  • New evidence of downstream utility (TAxn). We add a spoof-material classification experiment evaluated on held-out real images. TRACE augmentation achieves 75.51% accuracy, compared with 73.28% for standard augmentation, and also improves balanced accuracy and macro-F1. This provides task-specific evidence that controlled swaps produce useful training data (Section 4.3; Table 3; Appendix A.7).

These revisions connect TRACE’s methodological contribution with evidence of selective factor control, local ridge-structure preservation, and downstream augmentation utility, while making the scope of each evaluation explicit.

We appreciate your consideration of the revised manuscript and remain available to clarify any outstanding questions. Thank you.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.