{"id":"292e6318-4a27-4583-8747-5eba1ac04a94","arxiv_id":"2508.18031","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Applying four off-the-shelf GANs to 2D skull X-rays, the authors find FastCUT yields the best face reconstruction and retrieval results, but the evaluation rests on a small, self-collected dataset with a potentially leaky train-test split.","lead":"This paper applies standard GAN image-translation models to turn 2D facial X-rays into reconstructed face images, aiming to help identify crime and disaster victims. It reports that a lightweight model called FastCUT produces the most realistic faces and that retrieved matches often contain the correct identity, though the evidence rests on a very small, self-created dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random 80:20 pair split leaks identity across train/test, so FastCUT's identity preservation and retrieval scores may reflect memorization; the forensic-assistance conclusion is unsubstantiated.","rationale":"The paper is a proof-of-concept comparison of four GANs for 2D skull-to-face translation. Its central forensic-identification conclusion depends on two quantitative pillars: (a) identity preservation measured by ArcFace similarity between generated and ground-truth faces, and (b) retrieval recall@k against a face gallery. Both pillars require that test identities were never seen during training. Section 3 describes only a random 80:20 split of the 102 image pairs, which does not guarantee this; with 20 test pairs drawn from 102, most test identities will also have their other-view image in training. This is a concrete, checkable confound. If the model learned appearance from the training view of an identity, its generated test face can be a recombination of memorized identity features, which inflates ArcFace and retrieval scores. The paper's own caveat that FID/IS are unstable on 20 samples does not address leakage. The t-SNE visualization and Table 2b are therefore ambiguous. The proposed concrete test—re-splitting by 10 complete identities while preserving the 82/20 image counts—is exactly the experiment that would settle whether FastCUT actually generalizes. Because this decisive check is absent, the current evidence does not support the conclusion; I agree with the reader's REJECT verdict. Additional limitations (unverified gallery membership, no chance baseline, manual soft-tissue removal) are real but secondary; they do not change the verdict.","tokens_in":13930,"tokens_out":7155,"duration_ms":69569,"concrete_test":"Re-run the full pipeline with an identity-disjoint split that keeps the same sizes: randomly hold out 10 complete individuals (20 images, i.e., both frontal and lateral views) as the test set, and train on the remaining 41 individuals (82 images). Recompute Table 2 exactly: FID, IS, SSIM, LPIPS, ArcFace for all four models and recall@k/mAP@k for FastCUT with gallery sizes 100 and 200, verifying that the gallery actually contains the correct identity for each query. If FastCUT's ArcFace similarity drops substantially (e.g., below about 0.3) or recall@10 drops toward chance (10% for gallery=100), the original scores were inflated by identity leakage and the paper's conclusion fails; if the numbers remain essentially unchanged, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 6.1) is that generative models, especially FastCUT, can serve as an assisting tool for forensic craniofacial identification, supported by FID 63.65, SSIM 0.66, ArcFace similarity 0.5167, and high recall@k. For this claim to hold, the evaluation must measure generalization to previously unseen individuals. The paper's split in Section 3 ('Out of 102 pairs, 20 pairs are randomly selected for testing') is pair-level, not identity-level. Because each of the 51 volunteers contributes both a frontal and a lateral pair, a random selection of 20 pairs will, with high probability, include identities whose other view is in the training set: for a 20/102 sample, roughly 16 of the 20 test pairs are expected to come from identities also represented in training. No identity-disjoint split is described or enforced. Under these conditions, unpaired models such as CycleGAN, CUT, and FastCUT can reproduce identity-specific facial appearance learned from the training view of the same person, and the paired cGAN can likewise interpolate from a familiar identity. The reported ArcFace similarity and retrieval recall@k therefore do not establish that the models generate the face of an unseen skull; they may instead reflect memorization of training identities. The paper explicitly acknowledges FID/IS instability with 20 test samples (Section 5.2), but does not address this confounding overlap. Because the headline forensic-identification conclusion rests squarely on these identity-preservation and retrieval numbers, the claim is not supported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for craniofacial reconstruction from 2D X-ray skull images using four generative models (CycleGAN, cGAN, CUT, FastCUT). The authors extend the S2F dataset to 51 individuals (102 frontal/lateral image pairs), manually remove soft tissue from the X-ray images, train the generative models, and evaluate the generated faces using FID, IS, SSIM, LPIPS, and ArcFace similarity. They also propose a retrieval framework in which a generated face is used as a query against a gallery of real faces, reporting Recall@k and mAP@k for several backbones and gallery sizes. The paper concludes that FastCUT performs best and that these generative models can serve as an assisting tool for forensic craniofacial identification.","tokens_in":14118,"tokens_out":6227,"duration_ms":60613,"significance":"If the result held, the approach would be valuable because 2D X-rays are cheaper and more accessible than CT scans for forensic skull analysis. The paper's strengths include the construction of a paired skull–face dataset with ethical approval, a systematic comparison of four GAN-based methods, and a practical retrieval formulation with several feature extractors. However, the evaluation has a load-bearing flaw: the train/test split is by image pairs rather than by identity, so test identities almost certainly appear in training. In addition, the retrieval gallery is not specified to be disjoint from the dataset subjects, and no baseline is reported against direct skull-to-face matching. The central claim of forensic utility is therefore not currently supported.","major_comments":[{"comment":"The dataset is split into an 80:20 ratio, with 20 of 102 pairs randomly selected for testing. Because each of the 51 subjects contributes both a frontal and a lateral pair, this pair-level split almost certainly places the other view of a test identity in the training set. The reported ArcFace similarity (e.g., FastCUT: 0.5167) and Recall@k values in Table 2b may therefore measure the models' ability to reproduce identity-specific appearance seen during training rather than generalization to unseen skulls. Please re-run all experiments with an identity-disjoint split (both views of a subject in the same fold) and report the same metrics. This is the central issue for the conclusion in Section 6.1.","section":"Section 3, dataset split"},{"comment":"The face gallery is described only as 'collected from IITMandi student volunteers,' with no statement about whether it overlaps with the 51 dataset subjects. If the gallery contains face images of the identities used for testing (or of identities whose images appear in training), the retrieval results in Table 2b are inflated and cannot support the forensic identification claim. The authors must specify the gallery composition, the number of unique identities, and the degree of overlap with the training/test set, and should ideally evaluate with a public face database (e.g., LFW) for external validation.","section":"Section 4.6, retrieval gallery"},{"comment":"The conclusion states that 'these generative models can be used as an assisting tool for craniofacial identifications,' but no baseline comparison is provided. The paper should compare the proposed generative-retrieval pipeline against (a) direct skull-to-face matching using the cross-domain embedding method of ref. 13, (b) a simple nearest-neighbor baseline on skull image features, and (c) retrieval using the ground-truth face images as queries to calibrate an upper bound. Without these baselines, the added value of the generative reconstruction step is not demonstrated.","section":"Section 6.1, conclusion vs. baselines"},{"comment":"The manual soft-tissue removal procedure (Section 4.2) is not validated; if the masks do not accurately isolate skull structure, the 'skull' domain is ill-defined and all downstream results are questionable. Additionally, the authors acknowledge in Section 5.2 that FID and IS are 'not statistically stable' for 20 test samples, yet these metrics are used to rank methods. Please provide validation of the soft-tissue removal (e.g., inter-annotator agreement or comparison with CT-derived skulls) and report confidence intervals or bootstrap estimates for FID/IS, or relegate them to secondary consideration.","section":"Sections 4.2 and 5.2, validation of inputs and metric stability"}],"minor_comments":[{"comment":"The textual description of generators G_X and G_Y appears inconsistent with the equations and Figure 4: the text describes G_X as having a pretrained ResNet18 encoder and G_Y as fully trainable, but the loss definitions and figure assign G_Y to skull-to-face translation and G_X to face-to-skull translation. Please reconcile the notation.","section":"Section 4.3 and Figure 4"},{"comment":"The statement that FastCUT corresponds to lambda_X=10 and lambda_Y=0 should be checked against the original CUT paper; as written, it appears to contradict the standard FastCUT configuration (lambda_X=10, lambda_Y=0 in the original, but the surrounding text says lambda_X and lambda_Y are 1).","section":"Section 4.5, Eq. (11)"},{"comment":"The repeated 'Recall@k mAP@k' headers for each gallery size make the table difficult to parse; please use multi-level column headers or split the table into separate subtables for gallery sizes 100 and 200.","section":"Table 2b"},{"comment":"The phrase 'within the tok-k' is a typo and should read 'within the top-k'.","section":"Section 6.1"},{"comment":"The data availability statement says 'Our data will be made available on request,' but the face images used for training and testing are not shared due to privacy. Please clarify precisely which data (e.g., preprocessed masks, generated images, feature vectors) will be shared, as this affects reproducibility.","section":"Section 7, Data availability"}],"recommendation":"major_revision","confidential_remarks":"The identity leakage in the train/test split is the main technical concern and should be the first thing the authors address. The retrieval gallery needs explicit specification of its overlap with the dataset. The paper acknowledges its limitations, but the conclusion is stated too strongly relative to the current evidence. I would also note that the S2F dataset's face images are not public, so independent verification of the training data is difficult; a public gallery and a well-defined split would mitigate this."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper does something new and worth examining—first GAN-based skull-to-face translation from 2D X-rays, plus a modest extension of the authors' own S2F dataset—but the central quantitative evidence for the forensic claim is undermined by a train/test split that does not separate identities. I would send it to peer review with strong encouragement for revision, not desk reject it.\n\nWhat is actually useful: the task itself is legitimate and underexplored. Prior generative craniofacial work used CT or 3D data; 2D X-rays are cheaper and more accessible, so the application has practical value. The paper compares four standard GAN baselines (CycleGAN, cGAN, CUT, FastCUT), adds a retrieval module with four backbone networks, and extends the S2F dataset from 40 to 51 individuals. The authors also include failure cases and honestly acknowledge that FID and IS are unstable with 20 test samples. That is a reasonable proof-of-concept structure.\n\nThe soft spot is not minor. The dataset has 102 image pairs—each of the 51 subjects contributes a frontal and a lateral pair. The split in Section 3 is described as randomly selecting 20 pairs for testing, with no mention of an identity-disjoint split. With 20 test pairs out of 102, most test identities will also have their other view in the training set. For unpaired models like CycleGAN, CUT, and FastCUT, the face domain contains the same person's face during training, so the generator can learn to reproduce that identity's appearance rather than generalize from skull structure. ArcFace similarity of 0.52 and the high recall@k values likely reflect this memorization. The retrieval evaluation adds another problem: the gallery of 200 faces is from the same institution, but the paper never states whether the 51 volunteers overlap with the gallery, and there is no chance baseline or comparison against direct skull-to-face matching. With mAP below 10 in essentially all settings, the retrieval claim is weak even aside from the split issue.\n\nTo be fair, the paper is transparent about dataset size, demographic homogeneity, and lack of an external benchmark. The manual soft-tissue removal is described carefully, though not independently validated. But the identity leakage is a load-bearing flaw: the conclusion that these models can be an assisting tool for forensic identification rests almost entirely on identity-preservation and retrieval numbers that may simply reflect training identities.\n\nWho gets value from this? Researchers in forensic imaging or cross-modal translation might use the dataset extension or the baseline comparisons, but only after the evaluation is redone. The paper needs an identity-disjoint split, a specified gallery with no training overlap, chance-level baselines, and preferably more test samples. As it stands, the evidence supports feasibility at best, not forensic utility.\n\nMy recommendation: send to peer review, because the topic is real and the flaws are fixable. But a serious referee should require the re-evaluation before publication.","headline":"A genuinely new application of GANs to 2D X-ray craniofacial reconstruction, but the evaluation leaks identity across train/test, so the headline forensic-identification claim is not supported as stated.","tokens_in":14813,"tokens_out":2620,"would_cite":false,"duration_ms":28051,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A generative model can turn 2D skull X-rays into recognizable candidate faces, the paper argues.","keywords":["craniofacial reconstruction","forensic identification","generative adversarial networks","image-to-image translation","skull-to-face synthesis","2D X-ray","face retrieval","contrastive learning"],"falsifier":"Redo the training/test split so that all views of each person are kept together, then compare test ArcFace similarity and retrieval recall; if the scores fall to near chance, the current claims are inflated by identity leakage.","tokens_in":13600,"feed_emoji":"🦴","tokens_out":4607,"duration_ms":43451,"temperature":0.7,"pith_summary":"The paper tries to establish that generative models can reconstruct a face from a 2D X-ray of the skull, and that the best of them, FastCUT, preserves enough identity to act as an assisting tool in forensic craniofacial identification. This matters because standard forensic reconstruction relies on clay sculpting by experts or costly CT-based pipelines, whereas 2D X-rays are cheap and widely available. The paper reports that FastCUT beats CycleGAN, cGAN, and CUT on FID, IS, and ArcFace similarity, though cGAN is closer on perceptual similarity.","feed_headline":"AI model turns skull X-rays into recognizable faces","feed_subtitle":"Face-generating GANs could give forensics a low-cost route from 2D skull scans to candidate IDs.","key_machinery":"The load-bearing mechanism is FastCUT, a one-sided unpaired image-to-image translation model that combines an adversarial loss with a patch-wise contrastive (NCE) loss, matching each local patch of the generated face to the corresponding patch of the input skull while pushing other patches apart, plus an identity regularizer that penalizes changes to images already in the face domain. The pipeline also depends on a preprocessing step that manually removes soft tissue from X-rays so the input resembles a skull, and on the S2F paired dataset of frontal and lateral views. This machinery is what lets the model transfer structural details from skull to face without paired supervision.","core_discovery":"On its own terms, the paper's central claim is that skull-to-face translation can be done from ordinary 2D X-ray images rather than 3D CT data, and that the resulting faces are realistic and identifiable enough for forensic candidate retrieval. Using an extended S2F dataset of 51 individuals with frontal and lateral skull-face pairs, the authors compare four generative models and find FastCUT best: FID 63.65, SSIM 0.66, ArcFace similarity 0.5167. In retrieval tests against galleries of 100 and 200 real faces, correct identities appear in the top-k at high recall but with low mean average precision, so the authors position the framework as a candidate-shortlisting tool rather than a final identification system.","pith_inferences":["The reported identity-preservation numbers likely overstate generalization because a random pair split may put the same person's frontal and lateral views in both training and test; an identity-disjoint split is the natural next check.","If identity-disjoint evaluation confirms the result, 2D X-ray reconstruction could serve as a low-cost first-pass screening step to narrow missing-person databases before expensive DNA or CT analysis.","The high-recall/low-mAP pattern suggests the method is suited to candidate shortlisting rather than final identification, and retrieval ranking could be improved by training the generator with an explicit face-recognition loss.","Cross-population transfer remains untested because all subjects are young adults from one population; applying the model to skull X-rays from other age and ancestry groups would reveal whether the skull-to-face mapping generalizes."],"forward_implications":["If FastCUT's results hold, forensic teams could generate candidate faces from 2D skull X-rays in hours instead of weeks.","The same framework could be applied to lateral and frontal views separately, and the better-performing view could be used for retrieval.","The high recall but low mAP implies the system can narrow a missing-person database but should not be used alone for positive identification.","The comparison suggests lightweight one-sided models are more suitable than heavy cycle-consistency models for skull-to-face translation under limited data."],"supporting_citations":[{"why":"CycleGAN baseline for unpaired skull-to-face translation that the paper compares against.","marker":"[8]"},{"why":"The CUT and FastCUT models with patch-wise contrastive loss that the paper adapts and finds best.","marker":"[9]"},{"why":"The S2F benchmark dataset and cross-domain identity representation that the paper extends from 40 to 51 individuals.","marker":"[13]"},{"why":"cGAN/Pix2pix paired translation baseline used in the comparison.","marker":"[24]"},{"why":"ArcFace similarity used to measure identity preservation of generated faces.","marker":"[22]"},{"why":"FID metric used to measure generative quality.","marker":"[50]"},{"why":"LPIPS perceptual similarity metric used in the evaluation.","marker":"[21]"}],"fun_headline_variants":["GANs reconstruct faces from skull X-rays for forensics","Skull X-rays to faces: AI aids forensic ID","AI generates candidate faces from 2D skull scans","First use of 2D X-rays for AI face reconstruction","AI model shortlists suspects from skull X-rays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume that a random 80:20 split of the 102 skull-face pairs keeps test identities unseen, but because each of the 51 people contributes both a frontal and a lateral pair, the same person can appear in both training and test, so the identity scores may measure memorization rather than generalization.","fun_headline_variants_meta":{"raw":{"variants":["GANs reconstruct faces from skull X-rays for forensics","Skull X-rays to faces: AI aids forensic ID","AI generates candidate faces from 2D skull scans","First use of 2D X-rays for AI face reconstruction","AI model shortlists suspects from skull X-rays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1261,"prompt_tokens":943,"completion_tokens":318,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":239}},"tokens_in":559,"tokens_out":318,"duration_ms":3667,"temperature":1.0,"reasoning_tokens":239,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:58:14.276375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Redo the training/test split so that all views of each person are kept together, then compare test ArcFace similarity and retrieval recall; if the scores fall to near chance, the current claims are inflated by identity leakage.","supporting_citations":[{"cited_title":"A., Zhang, R","cited_arxiv_id":null,"evidence_quote":"The CUT and FastCUT models with patch-wise contrastive loss that the paper adapts and finds best."},{"cited_title":"& Efros, A","cited_arxiv_id":null,"evidence_quote":"cGAN/Pix2pix paired translation baseline used in the comparison."},{"cited_title":"& Zafeiriou, S","cited_arxiv_id":null,"evidence_quote":"ArcFace similarity used to measure identity preservation of generated faces."}],"review_version":2}