acceptodds
Under review as a conference paper at ICLR 2027

FAMILIAR: Conjuring the Many Faces of a Virtual Identity via Familiarized Occasion Priors

Abstract

Face synthesis conjures portraits of people who have never existed. Applications ask more of a virtual person than one portrait: a repertoire of appearances in which identity stays consistent. To create such repertoires, most existing generators treat each identity as a fixed point and add variation by a prescribed rule: perturbing its embedding, borrowing reference styles, or sampling predefined facial parameters. The form of an individual's variation is thus fixed by the method, and data at most fit its parameters. Cognitive research points to why this matters: familiarity with a face is built by learning how that face varies across encounters, and different faces vary along different dimensions. Following this insight, we introduce FAMILIAR, which learns person-specific variation instead of prescribing it. FAMILIAR represents a virtual identity as a distribution over photographic occasions. It first learns the occasion space through identity-conditioned image synthesis, without an attribute inventory, and then learns how that space is occupied for each individual: a conditional normalizing flow shared across identities predicts a full occasion distribution from a recognition embedding alone. A newly sampled identity therefore arrives with its own repertoire, without reference photographs. Occasion statistics differ more between people than within them, and with the generator fixed, another person's occasions lower recognition accuracy despite wider pose spread: useful variation is not interchangeable across identities. FAMILIAR therefore provides state-of-the-art synthetic data for face recognition and, by fixing a source occasion, de-identifies faces while preserving identity-irrelevant facial attributes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Face Synthesis, Identity-Conditioned Generation, Cognitive-Inspired Modeling, Synthetic Face Recognition, Face De-identification.

Reviews

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

This paper proposes FAMILIAR, inspired by research in cognitive science, which suggests that familiarity with a face stems from learning how it varies across repeated encounters, with different faces exhibiting distinct dimensions of variation. Unlike existing methods that rely on predefined scenarios for multi-dimensional face synthesis, FAMILIAR models a virtual identity as a "distribution of occasions." This approach maps an identity to a unique "occasion space," moving beyond the simple addition of arbitrary perturbations to a base identity. Extensive experiments validate two applications: first, sampling scenarios for new identities to achieve state-of-the-art (SOTA) performance in synthetic face recognition; and second, performing de-identification by swapping identities while keeping the source scenario constant, thereby minimizing identity leakage while preserving attribute utility comparable to the original images.

Critical strengths
  1. The motivation behind this paper is compelling. Drawing inspiration from cognitive science theories, it moves beyond fixed scenario settings. Instead, it posits that the form of the scenario itself should vary by individual and be learned rather than prescribed, which is an approach that yields valuable insights regarding identity and de-identification.
  2. The experimental comparisons in the paper are extensive and rigorous. For instance, with the generator fixed, comparisons are made against an empirical pool, unconditional flows, diagonal/low-rank Gaussians, mixture models, conditional flows, and mismatched-identity flows.
Critical weaknesses
  1. Evidence regarding the distribution of the generated data and the differences in how various identities undergo transformation remains unclear. The paper’s central argument is that each virtual identity possesses a unique set of appearance combinations, with distinct dimensions of variation for each identity. However, the paper fails to specify the exact content of these appearance combinations and lacks concrete statistical or qualitative results. Furthermore, the qualitative results presented appear to lack diversity, for instance, solely of frontal-view images with only slight angular deviations.
  2. Unequal training data confounds the state-of-the-art claim. FAMILIAR's generator is trained on WebFace-21M (20.7M images, 1.1M identities), whereas most baselines in Table 1 train on CASIA-WebFace (~0.5M images, ~10K identities), so the 0.6 – 1.1% margins cannot be attributed to the learned occasion prior rather than to roughly 40× more real training data. A FAMILIAR variant trained only on CASIA-WebFace is needed to support the claim.
Questions for authors

As weakness mentioned, I would like to provide qualitative results and scene distribution statistics to demonstrate diversity and that different identities have their own suitable scenarios. Also, please provide FAMILIAR test results on the same training dataset. If the authors provide the attribute statistics and interpretable-space evidence requested in W1, together with a same-data comparison for W2, I am open to raising my score.

4 comments
AuthorsOpenReview ↗

Response to Reviewer TMQJ [1/2]: content of occasions

We thank Reviewer TMQJ for the positive assessment and for concrete suggestions that led to new results.

We agree that what we call an occasion corresponds to the review's "scenario" or "scene". In FAMILIAR, an occasion is a learned code that captures the image-specific appearance of one photograph, without attribute labels (§3.1), and the revised §3.2 now makes its relation to identity explicit: "All identities share one occasion space, which spans photographic scenarios such as pose, expression, and illumination. Each identity has its own distribution over this space, so identities can differ in the occasions they favor."

W1 and Q1. Content of occasions and identities' own scenarios.

The revision adds the new Figure 5 to the main paper, and more results about attribute statistics and per-identity grids to Appendix B.13.

What occasions contain. An occasion is what the generator needs, besides identity, to reproduce a photograph (§3.1). Its content is interpretable. Linear probes on the occasion codes of held-out real photographs recover every measured attribute better than probes on the identity embedding, as shown in the table below:

AttributeOccasion codeIdentity embedding
Yaw0.980.03
Pitch0.760.21
Expression0.790.11
Age0.860.49
Brightness0.900.19
Lighting direction0.620.02
Color saturation0.870.23
Sharpness0.670.11

In the table, we report R2R^2 on 5,000 photographs of 200 unseen people, with probes fitted on 2,000 training identities. For expression, we report 7-class balanced accuracy (chance 0.14), and lighting direction is the left-right brightness asymmetry.

These results show that pose, expression, lighting, color, and sharpness are carried mainly by the occasion code (0.62 to 0.98, against at most 0.23 for the identity embedding). Only age is also predicted well from identity (0.49), as apparent age partly follows the person. Figure 5(a) in our revised paper gives the qualitative counterpart. Ten occasion codes selected from the identity-blind prior (one distribution shared by all identities), each rendered for four virtual identities, keep their pose (including large yaw), expression, and lighting down each column, while each row keeps its identity.

Identities' own occasions. Under the conditional prior, each identity has its own occasion statistics. In Figure 5(b) of the revised paper, among released identities selected for extreme statistics, one is predicted happy in all 50 of its samples and another in none. Across 2,000 identities, the table below tests whether each statistic is consistent within an identity:

StatisticConditionalIdentity-blindReal photos
Happy rate0.8620.1570.750
Expression entropy0.7300.1200.571
Mean yaw0.5010.0310.307
Yaw IQR0.5500.0020.382
Mean brightness0.7880.1260.745
Brightness SD0.167−0.0120.278
Mean color saturation0.7320.1730.744
Mean sharpness0.7810.1710.630
Mean age0.9880.4830.968

In the table, each entry is a split-half rr, the correlation across identities between each identity's statistic in two disjoint halves of its 20 images.

In the table above, each identity keeps its own tendencies under the conditional prior (rr 0.50 to 0.99, except brightness SD), at least as consistently as real people on seven of nine statistics. Under the identity-blind prior, rr stays below 0.18 except for age, which follows the rendered face.

These tendencies are also predictable for unseen people. Conditioned on five photographs of a person, the prior predicts the statistics of 20 other photographs of that person, as shown in the table below:

AttributeOwn personAnother personIdentity-blindCeiling
Happy rate0.570.070.120.91
Brightness0.380.020.310.90
Color saturation0.20−0.08−0.140.90
Sharpness0.180.040.070.87
Yaw0.100.00−0.060.59
Age0.940.380.780.99

In the table, the prior is conditioned on five photographs of the same person (Own person) or of a different person (Another person). We report Spearman ρ\rho over 200 unseen people, and the ceiling is the square root of the Spearman-Brown-corrected split-half reliability of the actual statistic. Renders use each person's own embedding, so age correlates under every prior.

These results show that the assigned tendencies follow the individual, even for people unseen in training. The person's own photographs give the highest correlation for every attribute, most clearly for happy rate (0.57 against at most 0.12).

Each identity's own occasions also serve it better. With the generator and identities fixed, another person's occasions lower accuracy from 94.29% to 92.86% despite wider pose spread (Table 2). For unseen real people, occasions predicted from one photograph raise rank-1 accuracy from 72.66% to 77.21%, against 73.78% with another person's occasions (Table 8).

AuthorsOpenReview ↗

Response to Reviewer TMQJ [2/2]: diversity of the generated data

W1 (continued). Diversity of the generated data.

The mostly frontal views in our figures reflect how we selected display examples. The generated data are as varied in pose as real photographs. In our released 0.5M dataset, 21.3% of images exceed an absolute yaw of 30° and 8.7% exceed 45°, against 21.5% and 8.3% for WebFace-21M photographs. We measured all datasets with the same estimators, and the table below compares their pose statistics:

DatasetAbs. yaw > 30°Abs. yaw > 45°Yaw IQRPitch IQR
Real (WebFace-21M)21.5%8.3%22.2°9.2°
FAMILIAR21.3%8.7%22.0°9.0°
MorphFace25.9%13.0%33.4°12.2°
SteerFace17.0%7.6%25.7°8.3°
UIFace17.4%6.8%24.6°8.3°
IDperturb16.4%6.0%25.7°7.8°
Vec2Face (HSFace10K)26.5%13.1%28.7°8.6°
Vec2Face+23.9%15.0%28.0°11.3°
DCFace13.6%5.3%21.0°7.6°
CemiFace12.8%4.2%22.2°6.8°
DigiFace21.8%7.9%28.2°8.8°

In the table, all synthetic datasets except HSFace10K and DigiFace-1M are the datasets of Table 1(b). From each, we use the first 2,000 identities with at least 20 usable images (up to 50 images each). For real photographs, we use 2,000 WebFace-21M identities with 20 images each. Image shares pool all of these images, whereas within-identity statistics, such as yaw and pitch IQR, use 20 random images per identity, averaged over identities.

These results show that FAMILIAR's pose statistics match those of real photographs. It is the only synthetic dataset within 4.8% (relative) of the real value on all four. Seven of the nine other datasets have a wider within-identity yaw IQR than real photographs (24.6° to 33.4°, against 22.2°), yet the five of them that we evaluated under one recognition recipe (MorphFace, SteerFace, UIFace, IDperturb, and Vec2Face+) reach 92.73% to 93.60% average accuracy at 0.5M (Table 1(b)), against 94.79% for FAMILIAR, so a wider pose spread alone does not make better training data.

FAMILIAR's expression distribution is also the closest to that of real photographs, as shown in the table below:

DatasetNeutralHappySadSurpriseFearDisgustAngerTV to real
Real (WebFace-21M)28.339.34.414.61.12.210.2—
FAMILIAR28.236.16.814.31.92.710.03.7
MorphFace24.338.27.810.63.43.012.68.9
SteerFace32.635.76.19.52.61.512.19.4
UIFace33.834.57.28.31.81.912.511.3
IDperturb32.633.17.08.72.01.914.712.3
Vec2Face (HSFace10K)41.338.63.28.90.20.87.013.0
Vec2Face+43.037.82.28.80.20.67.414.7
DCFace36.747.12.85.00.60.77.216.2
CemiFace40.644.02.15.00.80.57.017.1
DigiFace21.515.012.118.816.54.611.731.1

In the table, values are percentages of images, and TV is the total variation distance to the real distribution, computed before rounding. The revision adds both tables to Appendix B.13.

In the table above, every expression class of FAMILIAR is within 3.2 points of its real rate, and its TV of 3.7 is the lowest of all datasets (at least 8.9 for the others). Within each identity, FAMILIAR's spread is also close to that of real photographs, e.g., an expression entropy of 1.53 versus 1.54 bits and an age SD of 5.50 versus 5.51 years.

For qualitative evidence, Figure 5(a) in the main text shows estimated yaw from −40° to +35°, and the new Figure 7 (Appendix B.13) shows the range that the occasion space can render for nine released identities, with most occasion codes edited toward a target pose, expression, or lighting.

W2 and Q1. Unequal training data.

We agree. We retrained FAMILIAR on the CASIA-WebFace (CASIA) and compared it with the six CASIA-trained methods that are publicly available, under the same recognition recipe and protocol. FAMILIAR still leads on every benchmark:

MethodGenerator training dataLFWCFP-FPCPLFWAgeDBCALFWAvg.IJB-C 10−510^{-5}IJB-C 10−410^{-4}
DCFaceCASIA98.7587.6082.8389.2390.7989.8469.1682.11
CemiFaceCASIA99.0290.1987.7191.3192.1792.08N/AN/A
MorphFaceCASIA99.2893.9788.8992.2692.5293.3882.6488.76
SteerFaceCASIA98.9894.4488.7192.4392.9893.5181.5189.59
UIFaceCASIA99.2693.2388.2390.7492.2192.7381.4087.66
IDperturbCASIA99.0193.4488.4893.2893.5393.5585.2189.94
FAMILIARCASIA99.3294.9889.8293.7894.1194.4087.6291.90

In the table, every synthetic dataset has 0.5M images, and all rows use our IR-50/ArcFace recipe and the IJB-C template protocol. N/A: the recognizer trained on CemiFace's 0.5M dataset maps low-quality IJB-C frames to near-identical features, so TAR is near zero. The revision summarizes these results in Table 1(b) and gives the per-benchmark results and training details in Appendix B.11.

We would be glad to know whether the attribute statistics, the interpretable-space evidence (the new Figure 5), and this same-data comparison address your question.

Thanks for the detailed response. Most of my concerns have been addressed, so I will keep my score.

Thanks for your positive comments and insightful feedback, which have greatly strengthened our paper.

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

This paper investigates how to generate useful "intra-identity variations" for virtual faces. The proposed method first learns a compact representation of "occasions" (i.e., specific appearances or instances) through an "identity-conditioned latent rectified-flow" synthesis technique, utilizing frozen face-recognition embeddings to explicitly encode identity information. With the occasion encoder and the generator held fixed, a shared conditional normalizing flow is introduced to model the distribution of occasion encodings for a given identity embedding. This enables the system to sample diverse appearances for the same virtual identity during the generation phase, without requiring reference photographs. Furthermore, the generator supports "de-identification"—replacing the original identity information while preserving the source occasion encoding.

Validation efforts cover several areas, including face-recognition tasks based on synthetic data, comparisons of occasion priors using a fixed generator, and benchmarking for de-identification. The results demonstrate that, compared to identity-agnostic approaches, identity-conditioned occasion distributions yield data that is more valuable for training face-recognition models.

Critical strengths

1、Rather than simply viewing an increase in the degree of visual change as an inherently positive factor, this paper explores whether such change should instead be contingent upon the variable of "identity." Table 2 compares various "occasion priors"—including empirical, unconditional, Gaussian, mixture, and conditional-flow priors—while keeping the generator, identity information, and number of images constant. The results demonstrate that conditional-flow priors improve the average recognition accuracy from 93.18% to 94.29% compared to unconditional-flow priors.

2、This paper presents a range of analyses, including cross-photograph conditioning, ablation studies on temperature parameters and ranking, and scalability experiments conducted under fixed computational resource constraints. These analyses extend the value of this research beyond mere benchmark scores, offering a wealth of additional insights. Furthermore, the application of the same factorization scheme to both "recognition-oriented data generation" and "de-identification" highlights the versatility of this approach in practical scenarios. Although these two applications have yet to achieve fully disentangled representations, the complementary intervention strategies employed demonstrate that distinct conditioning channels serve different, specific functions within the model.

Critical weaknesses

1、The overall comparison fails to isolate improvements attributable specifically to the proposed "methodological" aspects. As noted in Appendix B.1, the methods under comparison differ across several dimensions, including generator training data, pre-trained components, downstream recognition recipes, and evaluation protocols. For instance, FAMILIAR utilizes WebFace-21M and DINOv3; consequently, even when synthetic datasets are matched in scale, there is no guarantee that the total supervision or computational resources invested are equivalent. Furthermore, it remains unclear whether there is any overlap between the identities used in the benchmarks and those present in the generator’s training data or the pre-training datasets. Therefore, Table 1 reflects the competitiveness of the overall system architecture rather than the superiority of the proposed method under strictly controlled, equivalent conditions.

2、The interpretation of "identity-specific variation" remains subject to potential confounding factors. In the "wrong-identity intervention" experiment, recognition performance declines (from 0.616 to 0.563) as the "consistency of the generated identity" decreases. This raises a question: is the performance drop due to the variation itself being "uninformative," or is it a result of "identity drift"—caused by an incompatible combination of identity and occasion? Consequently, while the empirical gains demonstrated by the method are compelling, any interpretation suggesting that "individual variations are non-interchangeable" (i.e., that they are not merely random fluctuations) remains, at this stage, suggestive rather than definitive.

Questions for authors

1、Can the authors conduct a comparison based on a unified experimental protocol, utilizing at least one accessible benchmark synthetic dataset? This should involve employing the same recognizer training recipe and the IJB-C evaluation criteria. If this is not feasible, which existing comparative studies come closest to controlling for these factors?

2、What checks have the authors performed regarding potential identity overlap or the presence of highly similar (near-duplicate) images between WebFace-21M and the evaluation benchmarks? Please clearly distinguish between instances that have been explicitly excluded and those that represent "unknown overlaps" within the external pre-training data.

3、When variables such as "identity consistency" and "image quality" are controlled for, do the advantages provided by the "conditional prior" still persist? Furthermore, do the "cross-photograph priors" or "support-mean priors"—mentioned in Table 3—continue to yield improvements in recognition performance when applied to the generation of synthetic datasets?

5 comments
AuthorsOpenReview ↗

Response to Reviewer frHo [1/3]: unified experimental protocol and training data

We thank Reviewer frHo for proposing the unified-protocol comparison and identity/quality-controlled tests.

W1 and Q1. Unified protocol.

We ran this comparison on all eight compared methods that are publicly available. We used the released datasets of five, regenerated UIFace and IDperturb with their official code, and reconstructed Arc2Face from its released model. FAMILIAR leads every method at both scales:

Method (0.5M images)LFWCFP-FPCPLFWAgeDBCALFWAvg.IJB-C 10−510^{-5}IJB-C 10−410^{-4}
DCFace98.7587.6082.8389.2390.7989.8469.1682.11
Arc2Face98.9792.7084.8691.2493.2792.2167.6281.83
CemiFace99.0290.1987.7191.3192.1792.08N/AN/A
MorphFace99.2893.9788.8992.2692.5293.3882.6488.76
UIFace99.2693.2388.2390.7492.2192.7381.4087.66
Vec2Face+99.2693.5987.9793.4593.7193.6079.8985.63
IDperturb99.0193.4488.4893.2893.5393.5585.2189.94
SteerFace98.9894.4488.7192.4392.9893.5181.5189.59
FAMILIAR99.4295.4189.8794.9994.2794.7987.8492.93
Method (1.0/1.2M images)LFWCFP-FPCPLFWAgeDBCALFWAvg.
Arc2Face99.1394.7385.9192.4093.0593.04
CemiFace99.0093.6989.5392.9093.2893.68
Vec2Face+99.3293.9489.1194.2794.3994.21
FAMILIAR99.4296.3391.3595.6094.8595.51

In both tables, Vec2Face+ at 1.0/1.2M uses its released set of 0.91M images. N/A: the recognizer trained on CemiFace's 0.5M dataset maps low-quality IJB-C frames to near-identical features, so TAR is near zero.

We also retrained FAMILIAR on the CASIA-WebFace (CASIA) training split alone. Under the same recipe, protocol, and 0.5M scale, it still leads all six evaluated methods whose generators are also trained on CASIA-WebFace:

MethodGenerator training dataLFWCFP-FPCPLFWAgeDBCALFWAvg.IJB-C 10−510^{-5}IJB-C 10−410^{-4}
DCFaceCASIA98.7587.6082.8389.2390.7989.8469.1682.11
CemiFaceCASIA99.0290.1987.7191.3192.1792.08N/AN/A
MorphFaceCASIA99.2893.9788.8992.2692.5293.3882.6488.76
SteerFaceCASIA98.9894.4488.7192.4392.9893.5181.5189.59
UIFaceCASIA99.2693.2388.2390.7492.2192.7381.4087.66
IDperturbCASIA99.0193.4488.4893.2893.5393.5585.2189.94
FAMILIARCASIA99.3294.9889.8293.7894.1194.4087.6291.90

The revision summarizes these results in Table 1(b) and gives the per-benchmark and 1.0/1.2M results and training details in Appendix B.11.

DINOv3, total supervision, and computation are not matched across methods, so §4.2 now calls Table 1(a) a system-level comparison. Table 2 isolates the method's effect. With pretrained components (ArcFace encoder, DINOv3, RAE decoder) fixed, the conditional prior adds 1.11 points of average accuracy and 2.24 points of IJB-C TAR at FAR 10−410^{-4} over the identity-blind prior. Appendices B.1 and B.10 (Table 19) report our training and generation costs.

Q2 and W1. Identity overlap.

WebFace-21M does contain some benchmark identities, but our audit finds no sign that FAMILIAR's synthetic data reproduce WebFace-21M people or that this overlap inflates our results. We excluded no benchmark identity ourselves (WebFace42M had already removed those overlapping the LFW family, FaceScrub, and IJB-C). Overlap with the pretraining data of the frozen ArcFace and of DINOv3 is unknown (Appendix B.1).

We first measured the overlap with a separate recognizer, using thresholds that only 0.1% of held-out real people or photographs exceed. Likely same-person matches in WebFace-21M occur mainly for AgeDB-30 (16.54% of images) and CFP-FP (10.95%). The only near-duplicate is one CALFW image (cosine of at least 0.95). However, our recognizers are trained only on synthetic images, so benchmark people would mainly affect our results if FAMILIAR reproduced WebFace-21M people. Our audit, which compares every synthetic identity and image with all of WebFace-21M (table in Response [3/3], item 3), flags about a tenth of the chance rate of identities and no generated image, and FAMILIAR's identities are farther from WebFace-21M than held-out real people are. We also checked the benchmarks directly. Benchmark faces are not closer to FAMILIAR's generated images than held-out real faces are (median highest cosine 0.316 versus 0.341 for the 0.5M dataset, and 0.330 versus 0.356 for the 1.2M dataset).

AuthorsOpenReview ↗

Response to Reviewer frHo [2/3]: identity consistency, image quality, and cross-photograph priors

W2 and Q3-1. Controlling identity consistency and image quality.

Yes, most of the advantage persists when the other priors reach the conditional prior's identity agreement at comparable image quality, so identity drift explains only part of the wrong-identity drop. In Table 2, identity agreement falls from 0.616 to 0.563 and accuracy from 94.29% to 92.86%. We ran an identity-matched comparison. With the generator, identities, and image count fixed, the identity-blind and wrong-identity priors draw 65 and 100 candidates per identity (the smallest tested budgets at which each reaches the conditional prior's identity agreement on validation identities) and keep the 50 with the highest conditioning-ArcFace similarity (the ranking step of the release policy, Appendix B.1). The table below reports identity agreement, image quality, and how well a recognizer trained on each dataset performs:

Occasion priorCandidatesIdentity agreementCR-FIQAAvg. accuracyIJB-C 10−410^{-4}
Conditional (Table 2)500.6162.37594.2992.07
Identity-blind, ranked650.6242.37193.5290.21
Wrong-identity, ranked1000.6272.37293.1589.81

In the table, identity agreement is the within-class AdaFace cosine as in Table 2, CR-FIQA is the mean image quality over all 500,000 images, the identity-blind prior is the unconditional flow of Table 2, and the ranked rows are means of three runs. The revision adds this table to Appendix B.12.

These results show that the conditional prior keeps most of its advantage once identity agreement is matched at nearly the same image quality. It keeps 0.77 of its 1.11-point lead over the identity-blind prior and 1.14 of its 1.43-point lead over the wrong-identity prior. Ranking raises the two priors' accuracy only from 93.18% and 92.86% (Table 2) to 93.52% and 93.15%. §4.3 now states the central finding operationally, and the Introduction and Conclusion follow: "With the generator fixed, useful variation is not interchangeable across identities: it helps recognition when it comes from the learned occasion distribution of the identity being rendered, whereas another person's occasions lower recognition accuracy despite wider pose spread" (a within-class yaw IQR of 22.4°, against 21.1° for the conditional prior in Table 2).

Q3-2. Cross-photograph and support-mean priors for dataset generation.

For the cross-photograph prior, yes. Most of the gain remains without the same photograph. We regenerated the Table 2 dataset (same 10K identities, generator, and sampling settings) with the "Other photo" prior of Table 3, which is trained with the identity embedding of another photograph of the same person, and conditioned it on each virtual identity. The table below compares it with the same-photo prior (the conditional flow of Table 2) and the identity-blind prior:

Occasion priorAvg. accuracyIJB-C 10−410^{-4}
Conditional, same photo94.2992.07
Conditional, other photo93.9091.12
Identity-blind93.1889.83

In the table, the same-photo and identity-blind rows are from Table 2, and the other-photo row is new.

In the table above, relative to the identity-blind prior, the other-photo prior raises average accuracy from 93.18% to 93.90% and IJB-C TAR from 89.83% to 91.12%, keeping 65% and 58% of the same-photo prior's gains (0.72 of 1.11 and 1.29 of 2.24 points). Like the likelihood gap in Table 3, the gain is associated mainly with the person.

We did not build virtual-identity datasets with the support-mean prior ("Mean of 5" in Table 3), because it is trained on normalized means of five embeddings, whereas a virtual identity is a single embedding drawn from a prior fitted to per-image embeddings (Appendix B.2). Where real photographs exist, this prior does improve recognition. In the enrollment experiment of Table 8, excerpted below, each of 200 unseen real people is enrolled with 1 or 8 real photographs, alone or with 16 images generated from the mean embedding of those photographs, using occasions from the prior named in each row:

Enrollment policy1 real support8 real supports
Real supports only72.6690.63
Unconditional occasions74.9191.38
Wrong-identity occasions73.7890.92
Conditional occasions77.2192.17

In the table, we report rank-1 accuracy (%) on 2,000 query photographs, and the conditional occasions come from the support-mean prior retrained for each support size (Appendix B.4).

In the table above, with one real photograph, conditional occasions raise rank-1 accuracy from 72.66% to 77.21%, against 74.91% with unconditional and 73.78% with wrong-identity occasions. With eight, the gain shrinks to 1.54 points (90.63% to 92.17%), and conditional occasions remain the best.

AuthorsOpenReview ↗

Response to Reviewer frHo [3/3]: ethics concerns

Thank you for raising this important concern. The revised Ethics Statement and Appendix B.14 now address these points.

1. Data provenance and permissions.

FAMILIAR is trained on WebFace-21M, the restored WebFace42M subset that the Arc2Face authors released on Hugging Face under the CC BY-NC-SA 4.0 license. This license permits non-commercial use and requires derivatives to be shared under the same license. We obtained the data from that release and use them only for this research. CASIA-WebFace trains only the real-data reference recognizer and the CASIA-WebFace variant of FAMILIAR (Appendix B.11), neither of which will be released. The recognition and de-identification benchmarks are used only for evaluation. We use all of these collections only for non-commercial research under their terms and redistribute none of them. The pretrained components, namely the frozen ArcFace encoder (InsightFace, non-commercial research), DINOv3 (DINOv3 License), and the RAE decoder, keep their own terms.

As the manuscript states, web-scale face datasets were collected without individual consent. For this reason, we release no real images and no real-face embeddings, and every released synthetic identity is audited against all training individuals.

2. Release conditions.

All releases are non-commercial and follow the licenses of the training data and the pretrained components. The frozen ArcFace encoder is not redistributed. We will release the code, the audited synthetic datasets with their synthetic identity embeddings, and the weights of the adapted decoder, occasion encoder, generator, and occasion prior. The identity prior is fitted to real identity embeddings and can propose identities close to training individuals, so it is shared only on request for non-commercial research under a signed data-use agreement. Each release will include a datasheet and a contact for removing any identity reported to resemble a real person.

3. Residual-identity risk assessment.

As you note, screening sampled identities against a finite gallery does not establish that generated images cannot resemble training individuals. Before release, we therefore audit every released identity against all WebFace-21M training identities, and every generated image against all WebFace-21M training images, with a separate recognizer. Each threshold is the 99.9th percentile of how close 20,555 held-out people (409,260 photographs) come to the training set, so chance alone flags 0.1% of held-out people. The table below gives the results:

Audited datasetIdentitiesFlaggedFlag rateMedian similarity
FAMILIAR 0.5M10,00010.010%0.430
FAMILIAR 1.2M24,00020.008%0.430
Held-out people20,555N/A0.1%0.454

In the table, Flagged counts identities above the identity threshold (N/A for held-out people, whose 0.1% rate holds by construction), and median similarity is the median over identities of each identity's highest cosine to any training identity. The revision adds this table to Appendix B.14.

These results show no sign of memorization. The audit flags about a tenth of the chance rate of identities and no generated image, and synthetic identities are slightly farther from the training set than held-out unseen people. Two procedural safeguards account for this. First, the identity prior is a smooth density fitted to the embeddings of the WebFace-21M training images and used only as a sampler. It models the population, whereas memorization would place samples near particular people, which is what the audit tests. Second, at sampling time a proposal is rejected if its cosine to any of 200,000 training identities reaches 0.3, where each training identity is represented by the photograph closest to its mean embedding. Because every identity is drawn independently and passes the same screen, the flag rate does not grow with the size of a released dataset. In the 10M dataset, four of 200,000 identities exceed the identity threshold (0.002%) and five others have an image above the image threshold. All flagged identities are removed from the released datasets, namely two across the 0.5M and 1.2M datasets (which share identities) and nine in the 10M dataset.

We would be grateful to know whether these replies address your questions and the ethics flag.

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

FAMILIAR studies how to generate diverse images of a virtual person while maintaining identity consistency. It jointly trains an occasion encoder and an identity conditioned image generator, where the occasion code captures image appearance without predefined attribute labels. After freezing these components, a shared conditional normalizing flow learns the occasion distribution pθ(c∣ϕ)p_\theta(c \mid \phi) from a recognition embedding. New identities can then be assigned their own distributions of occasion codes. The same generator also supports de-identification by retaining a source occasion code and replacing the identity. Validation includes synthetic face recognition benchmarks, controlled comparisons of occasion priors under a fixed generator, cross-photograph conditioning, covariance analyses, few shot enrollment, de-identification, and sampling policy ablations. The strongest evidence is that the conditional prior improves downstream recognition under controlled generation settings.

Critical strengths
  1. The central modeling contribution is useful and supported by a well controlled comparison. Learning an occasion space through synthesis and fitting a shared identity conditioned density provides a practical way to assign variation to newly sampled identities. Table 2 isolates the prior while holding the generator, identities, image count, and sampling policy fixed. The conditional flow improves average recognition accuracy from 93.18% to 94.29% and IJB-C TAR from 89.83% to 92.07% over the unconditional flow. Both have the same measured yaw spread. Conditional Gaussian and mixture alternatives provide meaningful intermediate baselines. This directly supports the contribution beyond simply using a stronger generator or generating more images.

  2. The paper makes a substantial effort to test alternative explanations for its results. Table 3 tests conditioning on other photographs of the same person, with the priors retrained for each conditioning construction. Table 6 includes appropriate finite sample covariance controls, and Table 5 separates the prior’s benefit from temperature mixing and candidate ranking. The few shot enrollment experiment is also valuable: conditional occasions improve one support recognition from 72.66% to 77.21%, compared with 74.91% using unconditional occasions, under an independent recognizer and disjoint support/query sets. De-identification provides a complementary application, and Table 10 directly tests the benefit of retaining the source occasion. Together, these experiments give the method stronger support than the main benchmark ranking alone.

Critical weaknesses
  1. The interpretation of person specific variation is stronger than what the interventions uniquely establish. The wrong identity experiment changes both the occasion distribution and its compatibility with the requested identity. In Table 2, within class identity consistency drops from 0.616 to 0.563 alongside recognition accuracy. The resulting accuracy loss can therefore partly reflect identity drift or less reliable synthetic labels. Appendix B.2 appropriately acknowledges that the swapped identity occasion pairs fall outside the generator’s training joint distribution.

The cross photograph and covariance controls provide meaningful evidence of person-associated structure, but that structure can also include residual identity information, acquisition patterns, or an identity dependent encoding of appearance. These experiments do not fully isolate characteristic nonidentity variation as the reason for the downstream gain. The well-supported conclusion is that matching the learned occasion distribution to identity improves synthetic data for this generator. The broader interpretation about an individual’s natural repertoire should be stated more cautiously.

  1. The external performance comparisons do not isolate the source of the reported superiority. Appendix B.1 states that Table 1 combines published results with different generator training data, pretrained components, downstream recognition backbones, losses, and schedules. FAMILIAR uses approximately 20.7M WebFace training images, while many baselines use CASIA-WebFace. Some IJB-C protocols also differ, including evaluation on still images versus the full template protocol. These differences limit how confidently the margins can be attributed to the proposed method.
Questions for authors
  1. Interpretation of the conditional-prior benefit. How much of the improvement should be attributed to more suitable intra-identity variation, versus improved identity consistency of the generated training data? The wrong-identity control changes both. Is there an existing analysis that separates these effects, or should the central conclusion be limited to the usefulness of matching the learned joint identity–occasion distribution? I am not requesting a broad new disentanglement study.

  2. Comparability of the main results. Which comparisons in Table 1, if any, use the same downstream backbone, loss, training schedule, and evaluation protocol as FAMILIAR? Please identify the directly comparable subset, particularly for IJB-C, and distinguish it from comparisons of published numbers. This clarification would help assess how much weight to place on the reported state-of-the-art margins without requiring every baseline generator to be retrained.

4 comments
AuthorsOpenReview ↗

Response to Reviewer bzzs [1/2]: interpretation of the conditional-prior benefit

We thank Reviewer bzzs for the positive assessment and for the precise formulation of what the wrong-identity control establishes.

W1 and Q1. Interpretation of the conditional-prior benefit.

We agree that the central conclusion should be limited. §4.3 now states it operationally, and the Introduction and Conclusion follow: "With the generator fixed, useful variation is not interchangeable across identities: it helps recognition when it comes from the learned occasion distribution of the identity being rendered, whereas another person's occasions lower recognition accuracy despite wider pose spread."

How much is due to identity consistency. Matching identity agreement closes only about a fifth to a third of the conditional prior's accuracy lead. To measure this, we ran an identity-matched comparison. With the generator, identities, and image count fixed, the identity-blind and wrong-identity priors draw 65 and 100 candidates per identity (the smallest tested budgets at which each reaches the conditional prior's identity agreement on validation identities) and keep the 50 with the highest conditioning-ArcFace similarity (the ranking of the release policy, Appendix B.1). We then trained a recognizer on each dataset. The table below reports the identity agreement, image quality, and recognition results of each dataset:

Occasion priorCandidatesIdentity agreementCR-FIQAAvg. accuracyIJB-C 10−410^{-4}
Conditional (Table 2)500.6162.37594.2992.07
Identity-blind, ranked650.6242.37193.5290.21
Wrong-identity, ranked1000.6272.37293.1589.81

In the table, identity agreement is the within-class AdaFace cosine as in Table 2, CR-FIQA is the mean image quality over all 500,000 images, the identity-blind prior is the unconditional flow of Table 2, and the ranked rows are the mean of three runs. The revision adds this table to Appendix B.12.

These results show that matching identity agreement does not close the gap. After ranking, both priors reach the conditional prior's identity agreement (0.624 and 0.627 versus 0.616) at nearly the same image quality, but they still train weaker recognizers. Without ranking, they reach an identity agreement of 0.588 and 0.563 and an accuracy of 93.18% and 92.86% (Table 2). Ranking therefore closes only 0.34 of the 1.11-point gap to the identity-blind prior and 0.29 of the 1.43-point gap to the wrong-identity prior, and the conditional prior keeps 0.77 and 1.14 points of its lead.

Existing analyses. Two existing results in the manuscript point the same way. Among the four conditional priors in Table 2, identity agreement varies by only 0.011 (0.605 to 0.616) while accuracy varies by 0.75 points and follows held-out NLL exactly. The diagonal Gaussian has higher agreement than the rank-32 and mixture priors yet trains the weakest recognizer. Candidate ranking raises the conditional prior's identity agreement by 0.032 (Table 5), more than its 0.028 lead over the identity-blind prior, yet raises accuracy by only 0.35 points.

Synthetic labels. Label noise does not explain the accuracy drop. We applied the label audit of Appendix B.9 to the three Table 2 datasets, as shown in the table below:

DatasetRejected same-label pairs
Conditional prior0.05%
Identity-blind prior0.14%
Wrong-identity prior0.21%
Real WebFace-21M images8.85%

In the table above, we report the share of same-label pairs that a separate audit recognizer rejects at FAR 10−310^{-3}. The revision adds this table to Appendix B.12. These results show that all three datasets reject far fewer same-label pairs than real images do, and the conditional-prior value matches the 0.05% of Appendix B.9. Even the wrong-identity dataset rejects only 0.21%, about a fortieth of the real-data rate, so label noise cannot explain the accuracy drop.

Person-associated structure. As suggested, §4.3 now states that the structure in Tables 3 and 6 can also include residual identity information, acquisition patterns, or identity-dependent encoding of appearance. The occasion code does carry some identity, but far less than the identity condition. Renders from the code alone match the source person in 15.19% of cases, against 96.31% with both inputs and 4.08% with neither (Table 10).

We answer W2 and Q2 in [2/2].

AuthorsOpenReview ↗

Response to Reviewer bzzs [2/2]: comparability of the main results

W2 and Q2. Comparability of the main results.

In Table 1(a), only the FAMILIAR row and the CASIA-WebFace real-data reference share one recipe and the IJB-C template protocol (§4.1, as stated in Appendix B.1). For all compared methods, we use their officially published results. We therefore revise the IJB-C sentence in §4.2, and the caption of Table 1 now states which values are published.

For a direct comparison, we applied our training recipe and the IJB-C template protocol to the eight methods with released code or datasets. We used the released datasets of five, regenerated UIFace and IDperturb with their official code, and reconstructed Arc2Face from its released model. FAMILIAR leads every compared method:

Method (0.5M images)LFWCFP-FPCPLFWAgeDBCALFWAvg.IJB-C 10−510^{-5}IJB-C 10−410^{-4}
DCFace98.7587.6082.8389.2390.7989.8469.1682.11
Arc2Face98.9792.7084.8691.2493.2792.2167.6281.83
CemiFace99.0290.1987.7191.3192.1792.08N/AN/A
MorphFace99.2893.9788.8992.2692.5293.3882.6488.76
UIFace99.2693.2388.2390.7492.2192.7381.4087.66
Vec2Face+99.2693.5987.9793.4593.7193.6079.8985.63
IDperturb99.0193.4488.4893.2893.5393.5585.2189.94
SteerFace98.9894.4488.7192.4392.9893.5181.5189.59
FAMILIAR99.4295.4189.8794.9994.2794.7987.8492.93
Method (1.0/1.2M images)LFWCFP-FPCPLFWAgeDBCALFWAvg.
Arc2Face99.1394.7385.9192.4093.0593.04
CemiFace99.0093.6989.5392.9093.2893.68
Vec2Face+99.3293.9489.1194.2794.3994.21
FAMILIAR99.4296.3391.3595.6094.8595.51

In both tables, Vec2Face+ at 1.0/1.2M uses its released set of 0.91M images, and the FAMILIAR rows are from Table 1(a). N/A: the recognizer trained on CemiFace's 0.5M dataset maps low-quality IJB-C frames to near-identical features, so TAR is near zero.

These results show that the lead does not come from the recipe. Using the same recipe and protocol, FAMILIAR leads by at least 1.19 points of average accuracy and 2.99 points of IJB-C TAR (FAR 10−410^{-4}) at 0.5M.

W2. Generator training data.

Retrained on the CASIA-WebFace training split alone, FAMILIAR still leads all six evaluated methods whose generators are also trained on CASIA-WebFace:

MethodGenerator training dataLFWCFP-FPCPLFWAgeDBCALFWAvg.IJB-C 10−510^{-5}IJB-C 10−410^{-4}
DCFaceCASIA-WebFace98.7587.6082.8389.2390.7989.8469.1682.11
CemiFaceCASIA-WebFace99.0290.1987.7191.3192.1792.08N/AN/A
MorphFaceCASIA-WebFace99.2893.9788.8992.2692.5293.3882.6488.76
SteerFaceCASIA-WebFace98.9894.4488.7192.4392.9893.5181.5189.59
UIFaceCASIA-WebFace99.2693.2388.2390.7492.2192.7381.4087.66
IDperturbCASIA-WebFace99.0193.4488.4893.2893.5393.5585.2189.94
FAMILIARCASIA-WebFace99.3294.9889.8293.7894.1194.4087.6291.90

In the table, every synthetic row uses the same training and evaluation protocol at 0.5M scale. These results show that FAMILIAR leads the strongest method IDperturb by 0.85 points of average accuracy. At 0.5M, it also exceeds Arc2Face and Vec2Face+, whose generators use more data. The revision summarizes the 0.5M and CASIA-WebFace results in Table 1(b) and gives the per-benchmark and 1.0/1.2M results and training details in Appendix B.11.

In addition, pretrained components also differ across methods, so §4.2 now calls Table 1(a) a system-level comparison. Table 2 isolates the method. With the generator and its pretrained components fixed, the conditional prior adds 1.11 points of average accuracy over the identity-blind prior.

We would be grateful to know whether the revision and the protocol-matched comparison results address your two questions.

Thanks to the authors' complementary experiments in the rebuttal. My concerns are resolved.

Based on the overall assessment and contribution of this work, I will maintain my original score.

Thanks for your positive comments and insightful feedback, which have greatly strengthened our paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response [1/2]

We sincerely thank all three reviewers for their constructive feedback and overall positive assessment.

We identify two primary concerns across the reviews:

  1. In Table 1, the generator training data and downstream face recognition recipes differ across the compared methods (all reviewers).

  2. It is unclear what the wrong-identity experiment demonstrates regarding identity consistency versus intra-identity variation (Reviewers bzzs and frHo).

Below, we first address these two shared concerns before responding to each reviewer individually.

1. Comparison with matched protocol and training data (all reviewers).

1-(a) Comparison with matched face recognition recipes. Under our unified recognition recipe and evaluation protocol, FAMILIAR still leads the evaluated methods, as shown in the two tables below:

Method (0.5M images)LFWCFP-FPCPLFWAgeDBCALFWAvg.IJB-C 10−510^{-5}IJB-C 10−410^{-4}
DCFace98.7587.6082.8389.2390.7989.8469.1682.11
Arc2Face98.9792.7084.8691.2493.2792.2167.6281.83
CemiFace99.0290.1987.7191.3192.1792.08N/AN/A
MorphFace99.2893.9788.8992.2692.5293.3882.6488.76
UIFace99.2693.2388.2390.7492.2192.7381.4087.66
Vec2Face+99.2693.5987.9793.4593.7193.6079.8985.63
IDperturb99.0193.4488.4893.2893.5393.5585.2189.94
SteerFace98.9894.4488.7192.4392.9893.5181.5189.59
FAMILIAR99.4295.4189.8794.9994.2794.7987.8492.93
Method (1.0/1.2M images)LFWCFP-FPCPLFWAgeDBCALFWAvg.
Arc2Face99.1394.7385.9192.4093.0593.04
CemiFace99.0093.6989.5392.9093.2893.68
Vec2Face+99.3293.9489.1194.2794.3994.21
FAMILIAR99.4296.3391.3595.6094.8595.51

In both tables, all rows use our IR-50/ArcFace recipe and the IJB-C template protocol. We evaluate all eight compared methods that are publicly available. We use the released datasets of five, regenerate UIFace and IDperturb with their official code, and reconstruct Arc2Face from its released model. N/A: the recognizer trained on CemiFace's 0.5M dataset maps low-quality IJB-C frames to near-identical features, so TAR is near zero.

These results show that FAMILIAR keeps its lead under one recipe and protocol, by at least 1.19 points of average accuracy and 2.99 points of IJB-C TAR (FAR 10−410^{-4}) at 0.5M images and by at least 1.30 points of average accuracy at 1.0/1.2M images.

1-(b) FAMILIAR retrained on CASIA-WebFace.

Retrained on the CASIA-WebFace training split alone, FAMILIAR still leads all six evaluated methods whose generators are trained on CASIA-WebFace, on every benchmark. The table below uses the same recipe, protocol, and 0.5M scale as in 1-(a):

MethodGenerator training dataLFWCFP-FPCPLFWAgeDBCALFWAvg.IJB-C 10−510^{-5}IJB-C 10−410^{-4}
DCFaceCASIA-WebFace98.7587.6082.8389.2390.7989.8469.1682.11
CemiFaceCASIA-WebFace99.0290.1987.7191.3192.1792.08N/AN/A
MorphFaceCASIA-WebFace99.2893.9788.8992.2692.5293.3882.6488.76
SteerFaceCASIA-WebFace98.9894.4488.7192.4392.9893.5181.5189.59
UIFaceCASIA-WebFace99.2693.2388.2390.7492.2192.7381.4087.66
IDperturbCASIA-WebFace99.0193.4488.4893.2893.5393.5585.2189.94
FAMILIARCASIA-WebFace99.3294.9889.8293.7894.1194.4087.6291.90

In the table above, FAMILIAR leads the strongest method IDperturb by 0.85 points of average accuracy. At the same 0.5M scale, it also exceeds Arc2Face and Vec2Face+ in 1-(a), whose generators use more data. The revision adds these results to the main text as Table 1(b), with details in Appendix B.11.

AuthorsOpenReview ↗

General response [2/2]

2. Suitable variation versus identity drift (bzzs-W1/Q1, frHo-W2/Q3).

We agree that the wrong-identity control in our original submission changes both the occasion distribution and its compatibility with the rendered identity. A new identity-matched comparison below shows that identity drift explains only about a fifth to a third of the accuracy gap. As Reviewer bzzs suggested, we also state our central conclusion operationally: "With the generator fixed, useful variation is not interchangeable across identities: it helps recognition when it comes from the learned occasion distribution of the identity being rendered, whereas another person's occasions lower recognition accuracy despite wider pose spread."

In the identity-matched comparison, the identity-blind and wrong-identity priors draw 65 and 100 candidates per identity (the smallest tested budgets at which each reaches the conditional prior's identity agreement on validation identities) and keep the 50 with the highest conditioning-ArcFace similarity, as in our release policy. The table below reports the identity agreement, image quality, and recognition results of each dataset:

Occasion priorCandidatesIdentity agreementCR-FIQAAvg. accuracyIJB-C 10−410^{-4}
Conditional (Table 2)500.6162.37594.2992.07
Identity-blind, ranked650.6242.37193.5290.21
Wrong-identity, ranked1000.6272.37293.1589.81

In this table, all datasets use the same generator and 10K identities. Identity agreement is the within-class AdaFace cosine as in Table 2, CR-FIQA is the mean image quality over all 500,000 images, the identity-blind prior is the unconditional flow of Table 2, and the ranked rows are the mean of three runs. The revision adds this table to Appendix B.12.

These results show that matching identity agreement does not close the gap. After ranking, both priors reach the conditional prior's identity agreement (0.624 and 0.627 versus 0.616) at nearly the same image quality, but they still train weaker recognizers (93.52% and 93.15% versus 94.29%). Without ranking, they reach 93.18% and 92.86% (Table 2), so ranking recovers only 0.34 and 0.29 points, and the conditional prior keeps 0.77 of its 1.11-point lead over the identity-blind prior and 1.14 of its 1.43-point lead over the wrong-identity prior. This indicates that most of the gain comes from the learned occasion distribution of the rendered identity rather than from identity consistency or image quality.

Summary of Revision

We made necessary revisions in the newly uploaded manuscript and marked all changes in blue:

  • Introduction, §4.3, Conclusion: operational central conclusion and scope of person-associated structure (bzzs, frHo).
  • §4.2, Table 1: system-level comparison, IJB-C protocols noted; Table 1(b): protocol-matched results at 0.5M and the CASIA-WebFace variant; Appendix B.11: per-benchmark and 1.0/1.2M results, training details, CemiFace IJB-C diagnosis (all reviewers).
  • Appendix B.12: identity-matched comparison, label audit, image quality, cross-photograph generation (bzzs, frHo).
  • §4.3 (new Figure 5, with old Figure 5 now Figure 6), Appendix B.13: occasion content, identity differences, attribute statistics, per-identity grids (TMQJ).
  • Ethics and Reproducibility statements, Appendix B.14: data provenance, release conditions, residual-identity audit, misuse, benchmark overlap (frHo).
  • Appendix B.1: directly comparable subset, exclusions versus unknown overlaps, screening gallery, data availability, released datasets (bzzs, frHo).
  • §3.2: one shared occasion space with per-identity distributions (TMQJ).
  • Appendix C: provenance of the new figures.

In the responses to each reviewer below, W and Q refer to the Weaknesses and Questions in that review, respectively.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.