{"id":"9e815e4e-68e0-4c10-aa9f-6532567a1ef3","arxiv_id":"2509.06282","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Selfie images can be used to predict skin hydration and water loss with R2 up to about 0.35, using a new dataset of 336 panelists and an adapter-based vision transformer.","lead":"The authors collected a dataset that links 37-point measurements of skin hydration and water loss to smartphone selfies and clinical facial photos, and trained an adapted vision transformer to predict these values from skin patches. The model beats standard image networks, but the correlations are modest and it is unclear how much of the prediction comes from the image itself versus the facial location being measured.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing per-position mean baseline makes reported R2 gains uninterpretable as image-driven assessment.","rationale":"The reader's weakest_assumption identifies exactly this concern, and I agree. The paper's own data show strong per-region label variation (Fig. 8), and the largest ablation jump comes from explicit position IDs (Table 2). The ViT-B comparison does not resolve the issue because standard ViT can infer facial location from patch context; the missing control is the position-only predictor. Anchor-estimation noise (average error 0.77–0.78 sticker radii, §4.4) further weakens the image–label alignment, making it even more important to know how much signal comes from the position embedding rather than from skin pixels. The conclusion's admitted limitations (lighting diversity, weaker SH) do not address this gap. No confidence intervals and lack of code/data release are secondary but reinforce the conditional stance. If the position-only baseline matches Config C, the central claim collapses; if it is far lower, the paper's image-driven story is supported. Because this control is easy to run and the paper is otherwise plausible, the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":23374,"tokens_out":6894,"duration_ms":83591,"concrete_test":"On the same train/test split as Table 1, compute a position-only baseline: for each anchor ID d in {1..37}, assign every test patch with that ID the training-set mean of the corresponding SH/TEWL label (or a per-ID median/linear model on ID alone). Report R2 and MAE for TEWL and SH on both Selfie and VISIA test sets, including Many/Med/Few subgroups. If this baseline's R2 is within ~0.02 of Config C (or final, Config E), the position prior dominates and the image-texture contribution is unproven; if it is substantially lower, the concern is mitigated. Ideally also report the same baseline for Config C trained with position IDs randomly permuted, to confirm the model is not using position only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is the absence of a per-position mean-only baseline. In §3.3, Position Adapters take the 1-of-37 anchor ID as a one-hot vector and prepend a position prompt at every transformer layer; Table 2 shows the largest single R2 jump when these are added (Config C: Selfie TEWL 0.210→0.253, SH 0.109→0.190; VISIA TEWL 0.311→0.320, SH 0.200→0.231). Since each facial region has its own SH/TEWL distribution (Fig. 8), a trivial predictor that outputs the training-set mean for each anchor ID could plausibly reproduce much of this R2. If so, the model is memorizing regional averages rather than reading skin appearance, and the central claim of remote, image-driven assessment is not supported. The reported ViT-B baselines do not settle this, because ViT-B can also infer facial location from patch context; the missing control is the position-only predictor. Anchor-estimation error (0.77–0.78 sticker radii, §4.4) further loosens the image–label coupling, making it essential to know how much signal is carried by the position embedding rather than by skin pixels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a complete pipeline for estimating facial skin hydration (SH) and trans-epidermal water loss (TEWL) from smartphone and VISIA facial images. Data were collected from 336 Chinese panelists at 37 facial anchors using a Corneometer and a VapoMeter; anchors are localized in non-sticker images by a DAD-3DHeads + PointNet pipeline. The proposed Skin-PA ViT uses a frozen ImageNet-pretrained ViT with texture adapters, a prior texture module with frequency filtering, position adapters fed by 1-of-37 anchor IDs, lighting augmentation, and a facial-symmetry contrastive loss. On subject-disjoint test splits, the method reports R2 = 0.264 (TEWL) and 0.205 (SH) on Selfie images and R2 = 0.348 and 0.263 on VISIA images, outperforming several CNN and ViT baselines. Ablations attribute the gains incrementally to the proposed components.","tokens_in":23616,"tokens_out":4787,"duration_ms":49657,"significance":"If the reported predictive power comes from skin appearance rather than from positional priors, this would be a useful step toward accessible, image-based skin-barrier assessment. Strengths of the manuscript include a subject-disjoint train/test split stratified by skin tone, evaluation on two capture modalities, multiple strong baselines, and an ablation study that is reported transparently. The work is also honest about dataset limitations and the weaker performance on selfie images. However, the central claim that SH and TEWL are estimable from RGB patches is not yet established: the position-ID input is a major confound, and the reported anchor-estimation error loosens the image-label coupling. The missing per-position mean-only baseline is therefore load-bearing and must be addressed before the image-driven claim is convincing.","major_comments":[{"comment":"The Position Adapters take the 1-of-37 anchor ID as a one-hot input and prepend a position prompt at every transformer layer. Table 2 shows the largest single R2 jump when these adapters are added (Config B to C: Selfie TEWL 0.210→0.253, SH 0.109→0.190; VISIA TEWL 0.311→0.320, SH 0.200→0.231). Figure 8 shows that the per-position mean SH/TEWL values vary substantially across the face. A trivial predictor that outputs the training-set mean for each anchor ID could plausibly reproduce much of this gain. Because no such per-position mean baseline is reported, the reported R2 values are uninterpretable as evidence that the model reads skin appearance rather than memorizing regional averages. Please add a per-position mean-only baseline (and ideally a position-ID-only model with no image input) for both datasets and for the shot-stratified MAE columns. This is necessary to support the central","section":"§3.3, §4.7, Table 2"},{"comment":"The anchor estimator reports average error rates of 0.77 and 0.78 sticker radii for Selfie and VISIA images, and the paper states these are acceptable because the error is within the sticker radius. However, the SH/TEWL ground-truth labels are point measurements taken at the sticker centroids, while the test-time patches are cropped around the predicted anchors. An offset of ~0.78 radii can place the cropped patch mostly outside the region that was physically measured, weakening the image-label alignment. The paper does not quantify the effect of this misalignment on the reported R2. Please report the evaluation using ground-truth sticker centroids on the test set (or a sticker-jitter sensitivity analysis) to show how much performance degrades when anchor error is removed, and discuss the implications for the reliability of the learned image-label mapping.","section":"§4.4"},{"comment":"All regression metrics are computed over cropped patches, but patches are highly correlated within a subject: the same subject appears in multiple views, lighting conditions, and nearby facial anchors. The differences between the proposed method and the strongest baselines are small in several cases (e.g., Selfie TEWL R2 0.264 vs. 0.236 for Swin-B; VISIA TEWL R2 0.348 vs. 0.318 for ViT-B). Without subject-level clustering, confidence intervals, or a paired test, it is unclear whether these differences are statistically meaningful. Please report cluster-robust standard errors or subject-level bootstrap confidence intervals for the headline R2 and MAE numbers.","section":"§4.2, Table 1"}],"minor_comments":[{"comment":"The paper defines M=37 anchor points per facial image, but §4.1 states that 19 skin patches are cropped from each facial image. This inconsistency affects the dataset-size calculation and the meaning of the position-ID input. Please clarify which anchors are visible or used per image/angle and correct the notation.","section":"§3.2 vs. §4.1"},{"comment":"The Position Adapter input is described as a one-hot vector of anchor IDs, but the relationship between the 37 global anchor IDs and the 19 patches per image is not specified. At test time, is the position ID the global facial anchor ID or an image-relative index? Please define the encoding precisely.","section":"§3.3"},{"comment":"The objective function uses Lcon in Eq. (10) while Eq. (9) defines ℓcon; please unify the notation. Also clarify the construction of the negative set Z in Eq. (9): are negatives drawn from other positions, other subjects, and/or other images in the batch? This affects the interpretation of the contrastive loss.","section":"§3.5, Eq. (10)"},{"comment":"The text says the final Config E achieves superior performance, but for VISIA TEWL the MAE (All) increases from 2.16 (Config D) to 2.19 (Config E). Please qualify the claim or discuss why the symmetric contrastive loss trades overall MAE for better few-shot MAE.","section":"Table 2"},{"comment":"The lighting generalization results are presented only as a figure with average R2 improvements. Please report the numeric values, including the without-augmentation baseline for each leave-one-lighting-out direction, so the gains are quantitatively assessable.","section":"§4.6, Fig. 15"},{"comment":"Minor language issues: 'par with' in §4.6, 'it's' in the Acknowledgment, and the duplicated caption text in Fig. 2. These do not affect the technical content.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the missing per-position mean-only baseline. If that baseline reproduces much of the reported R2, the central claim should be reframed as regional-prior prediction rather than image-texture-driven assessment. The anchor-error analysis is also important because it bears directly on whether the model is learning a meaningful image-label mapping. The dataset is not publicly available and no code is released; for a systems paper this is understandable given PII constraints, but it further raises the bar for reporting these control analyses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read the paper alongside your stress test, and I think the position-ID concern is real and load-bearing. The paper's own ablation shows the largest single jump in Table 2 when Position Adapters are added (Config C), and the model feeds a 1-of-37 anchor ID into every layer. Without a per-position mean-only baseline, we don't know whether the model is reading skin texture or just memorizing regional averages. Fig. 8 shows real regional variation, so this isn't pedantic. I'd want to see that control before believing the remote-assessment claim.\n\nThat said, the paper has real contributions. The 37-point SH/TEWL dataset spanning selfie and VISIA images is new and useful, even if it's not released due to PII. The systematic pipeline—anchor estimation, patch cropping, model, heatmap—is described well enough to reproduce. The evaluation uses subject-disjoint splits and standard devices, and the paper honestly reports the modest R2 values (0.264 TEWL / 0.205 SH on Selfie) plus the limitations (limited lighting diversity, weaker SH signal). No circularity; labels come from physical devices.\n\nMinor issues: anchor error of 0.77–0.78 sticker radii is accepted without quantifying its effect on label noise; some ablation differences are within rounding distance; no confidence intervals anywhere. The dataset isn't available, so external verification is hard.\n\nThe per-position mean baseline is the one that matters. If the model collapses to regional averages, the method is not image-driven assessment. If the baseline is added and the image-texture contribution remains, the paper is solid. I'd recommend sending it to peer review with that control as a required addition. The work is relevant to anyone building image-based physiological measurement, and the confound discussion is a useful cautionary tale for the field.","headline":"A genuinely new dataset for selfie-based SH/TEWL, but the missing per-position mean baseline leaves the central claim unproven—still worth refereeing.","tokens_in":24162,"tokens_out":3518,"would_cite":false,"duration_ms":32978,"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":"Skin hydration and water loss can be estimated from selfie images","keywords":["skin hydration estimation","transepidermal water loss","selfie image analysis","vision transformer adapters","imbalanced regression","contrastive learning","skin barrier function","facial skin assessment"],"falsifier":"A position-only control: train Skin-PA ViT with the 37-position one-hot ID but a blank or averaged image (or otherwise ablate image information). If it reaches R2 near 0.264 for TEWL or 0.205 for SH on the Selfie test set, the central remote-assessment claim—that skin appearance is being read—would fail.","tokens_in":23244,"feed_emoji":"📱","tokens_out":7726,"duration_ms":80321,"temperature":0.7,"pith_summary":"Dermatologists measure skin barrier function with point instruments—a Corneometer for stratum corneum hydration and a VapoMeter for trans-epidermal water loss—but these devices are expensive and clinic-bound. This paper claims that both quantities can instead be estimated from ordinary RGB selfie images, making full-face skin-barrier assessment as simple as taking a photo. To establish this, the authors built a dataset of 336 panelists measured at 37 standardized facial anchors and paired with smartphone and controlled VISIA images, and trained Skin-PA ViT, a frozen Vision Transformer augmented with texture and position adapters plus a symmetry-based contrastive regularizer. On held-out test patches, the model reaches R2 = 0.264 for TEWL and R2 = 0.205 for SH from Selfie images, outperforming standard CNN and ViT baselines, with the clearest gains on rare value ranges. If the claim holds, remote, low-cost skin-barrier monitoring—including heatmaps of where the barrier is weak—becomes feasible for general users.","feed_headline":"Selfie AI maps skin hydration and water loss without probes","feed_subtitle":"A vision transformer trained on 37 facial measurement points reads skin barrier health from phone photos.","key_machinery":"The load-bearing object is the Skin-Prior Adaptive Vision Transformer (Skin-PA ViT), a regression head built on a frozen ImageNet-pretrained ViT-B. Instead of tuning the backbone, the method trains three small components: a Prior Texture Module (a small CNN) whose output is turned into prompt tokens by Texture Adapters, and Position Adapters that convert the facial anchor ID (1 of 37) into prompt tokens prepended to the input sequence of every Transformer layer. The prompt tokens let each layer use both image texture and knowledge of where on the face the patch sits. The input patch is concatenated with a band-pass-filtered version of itself in the Fourier domain, which keeps mid-range spati","core_discovery":"The paper's central claim is that hydration (SH) and trans-epidermal water loss (TEWL), two point-measured markers of skin barrier function, are readable from small RGB patches of facial skin. The authors collected paired instrument measurements at 37 facial anchors and facial images from 336 panelists, then trained a regression model that predicts the two values from each cropped patch. Their model, Skin-PA ViT, keeps an ImageNet-pretrained Vision Transformer frozen and learns only lightweight adapters that inject high-frequency texture tokens and a one-hot facial-position token into every Transformer layer; a band-pass filtered version of the patch is used to highlight wrinkles and pores,","pith_inferences":["The paper leaves implicit that a position-only control—training the same architecture with the 1-of-37 position ID but with image content ablated—would quantify how much of the R2 is regional prior rather than skin appearance; that experiment is a direct test a reader could run.","Because the dataset is drawn from 336 Chinese panelists with a skin-tone-based train/test split, the method's cross-ethnicity generalizability is untested; a natural extension is to evaluate on panels spanning broader ITA values and device types.","The anchor estimator's average error of 0.77–0.78 sticker radii means training labels are spatially noisy; one testable extension is to train with label-smoothing or multiple random crops per anchor and measure whether accuracy improves.","A consumer version would need to handle occlusions, extreme poses, and makeup; since the model is evaluated on multi-angle, multi-lighting images, a natural stress test is to probe those real-world distortions systematically."],"forward_implications":["If the method generalizes beyond the collected population, an ordinary smartphone photo can produce full-face SH/TEWL heatmaps, removing the need for a clinic visit for routine barrier-function checks.","The position-adapter design shows that injecting explicit anatomical region information helps a small-data regression model, suggesting a recipe for other per-region physiological estimates from images.","The symmetry-based contrastive loss and lighting augmentation both reduce error on rare value ranges, pointing to a general strategy for imbalanced regression where labels are distributed unevenly across spatial regions.","The data-collection and anchor-estimation pipeline provides a reproducible protocol for building future skin-health image datasets with point-instrument labels.","Because SH prediction trails TEWL, the paper's own framing implies that adding non-RGB cues (for example infrared or other modalities) would be the natural next step for hydration."],"supporting_citations":[{"why":"Supplies the frozen ViT-B backbone used by all models and is the main baseline to beat.","marker":"[7]"},{"why":"Closest prior work relating TEWL to skin texture from cheek images; the paper extends it to full-face selfie assessment and cites it for SH's weaker texture link.","marker":"[6]"},{"why":"Supplies the facial landmarks that anchor estimation starts from.","marker":"[49]"},{"why":"Maps landmark sets to the 37 measurement-anchor coordinates used to crop training patches.","marker":"[50]"},{"why":"Supervised contrastive loss formulation adapted as the facial symmetry-based regularizer.","marker":"[52]"},{"why":"Supplies the regional SH/TEWL variation and skin-tone/hydration correlation that motivate position adapters, lighting augmentation, and full-face assessment.","marker":"[3]"},{"why":"Visual prompt tuning supplies the prompt-token mechanism that the Texture and Position Adapters build on.","marker":"[9]"}],"fun_headline_variants":["Selfie AI predicts skin hydration and water loss","AI reads skin barrier from selfies, no probes needed","Vision transformer assesses skin from facial photos","Smartphone selfie reveals skin hydration and TEWL","Selfie-based AI estimates skin barrier health without instruments"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a skin patch's appearance, not its labeled position on the face, is what carries the SH/TEWL signal; the model is given both, so the claim stands or falls on whether image texture is genuinely predictive once position is controlled.","fun_headline_variants_meta":{"raw":{"variants":["Selfie AI predicts skin hydration and water loss","AI reads skin barrier from selfies, no probes needed","Vision transformer assesses skin from facial photos","Smartphone selfie reveals skin hydration and TEWL","Selfie-based AI estimates skin barrier health without instruments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1325,"prompt_tokens":738,"completion_tokens":587,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":513}},"tokens_in":482,"tokens_out":587,"duration_ms":6470,"temperature":1.0,"reasoning_tokens":513,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:50:52.648625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A position-only control: train Skin-PA ViT with the 37-position one-hot ID but a blank or averaged image (or otherwise ablate image information). If it reaches R2 near 0.264 for TEWL or 0.205 for SH on the Selfie test set, the central remote-assessment claim—that skin appearance is being read—would fail.","supporting_citations":[{"cited_title":"Assessment of skin barrier function using skin images with topological data analysis,","cited_arxiv_id":null,"evidence_quote":"Closest prior work relating TEWL to skin texture from cheek images; the paper extends it to full-face selfie assessment and cites it for SH's weaker texture link."},{"cited_title":"Dad-3dheads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image,","cited_arxiv_id":null,"evidence_quote":"Supplies the facial landmarks that anchor estimation starts from."},{"cited_title":"Pointnet: Deep learning on point sets for 3d classification and segmentation,","cited_arxiv_id":null,"evidence_quote":"Maps landmark sets to the 37 measurement-anchor coordinates used to crop training patches."},{"cited_title":"Supervised contrastive learn- ing,","cited_arxiv_id":null,"evidence_quote":"Supervised contrastive loss formulation adapted as the facial symmetry-based regularizer."},{"cited_title":"Voegeli, J","cited_arxiv_id":null,"evidence_quote":"Supplies the regional SH/TEWL variation and skin-tone/hydration correlation that motivate position adapters, lighting augmentation, and full-face assessment."},{"cited_title":"Visual prompt tuning,","cited_arxiv_id":null,"evidence_quote":"Visual prompt tuning supplies the prompt-token mechanism that the Texture and Position Adapters build on."}],"review_version":1}