{"id":"db8e6ac3-c218-4f79-b622-4140187af2b8","arxiv_id":"2509.05078","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"A network that feeds three pooled multi-scale CNN features through a two-block transformer reaches PC 0.9187 on SCUT-FBP5500, edging out the cited R3CNN baseline.","lead":"This paper describes a hybrid CNN-transformer network that combines multi-scale facial features and predicts beauty scores, reporting a 0.9187 Pearson correlation on the SCUT-FBP5500 benchmark. A generalist may care as a compact example of attention across image scales, but the improvement over previous work is small and only shown on one data split.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on Table 2 comparisons that are neither protocol-matched nor complete; the 0.0045 PC margin over a 2019 baseline is within plausible run-to-run variance and recent transformer FBP methods are omitted.","rationale":"The paper is a straightforward empirical architecture study, and the proposed SIT design—multi-scale CNN features fed as a sequence into a small transformer—is internally coherent. No fundamental mathematical or architectural error is apparent; the equations and algorithms are plausible. The load-bearing issue is entirely about the empirical evidence for the headline SOTA claim. The reader's weakest_assumption focused on the single-fold, no-variance protocol and the possibility of metric/preprocessing mismatch; I agree that this is a serious concern. I extend it with two concrete observations that make the concern even sharper. First, Table 2 is not only protocol-uncontrolled but also incomplete: the reference list includes several recent transformer-based FBP papers by the same author ([13], [30]), and a state-of-the-art claim must compare against them. Without knowing their scores, the 0.9187 figure cannot be certified as SOTA. Second, there is an internal inconsistency between the abstract's 'both correlation and error metrics' and the MAE column in Table 2, where R3CNN is better. This does not by itself falsify the PC/RMSE claim, but it weakens the paper's consistency and suggests the comparison narrative is not carefully constructed. The proposed concrete test—a multi-seed, protocol-matched comparison against an expanded baseline set—would settle whether the SOTA claim holds. Given the severity of these empirical gaps, the reader's CONDITIONAL verdict is appropriate; I would not move it to ACCEPT or REJECT based on the current evidence. The architecture may well be valid, and the claims could survive revision, but the evidence as presented is insufficient to establish a new state-of-the-art.","tokens_in":9444,"tokens_out":4883,"duration_ms":55537,"concrete_test":"Release or re-implement the SIT code and run it on SCUT-FBP5500 fold 3 with the exact preprocessing in Section 3.1, alongside all Table 2 baselines and the omitted recent methods (at minimum [13] Global Context ViT and [30] Mamba-CNN), using at least 5 seeds. Report mean ± std for PC, RMSE, MAE. If any omitted method's mean PC >= 0.9187, or SIT's mean PC advantage over R3CNN is less than the pooled std, then the 'new state-of-the-art' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim (PC 0.9187, RMSE 0.2760) is supported only by Table 2, which compares against a small, largely pre-2020 baseline set. The paper's own Section 5.1 admits single-fold reporting; no variance or seed information is given. The margin over R3CNN (0.0045 PC, 0.004 RMSE) is smaller than typical run-to-run variation for this dataset, so even a correct re-run may not reproduce the ordering. Table 2 also omits recent transformer-based FBP methods, including the author's own [13] and [30], so 'new state-of-the-art' is asserted without evidence that these methods are worse. Additionally, the Abstract/Section 1 claim of outperforming 'both correlation and error metrics' is contradicted by Table 2, where R3CNN's MAE (0.2120) is lower than SIT's (0.2180). These issues undermine the empirical foundation of the claim, independent of the architecture's internal soundness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Scale-Interaction Transformer (SIT), a hybrid CNN-transformer for facial beauty prediction on SCUT-FBP5500. The architecture extracts a MobileNetV2 feature map, applies three parallel convolutions (1×1, 3×3, 5×5), pools each with GAP and GMP, stacks the resulting three 128-dimensional vectors into a sequence, and processes this sequence with a two-block transformer encoder before a regression head. The reported results are PC 0.9187, MAE 0.2180, RMSE 0.2760 on fold 3, which the authors claim is a new state-of-the-art, surpassing R3CNN (PC 0.9142, RMSE 0.2800) and prior methods. Ablations show progressive improvement from the CNN baseline to the full model, supporting the contribution of the multi-scale and transformer modules. The paper also includes a limitations section acknowledging single-fold evaluation.","tokens_in":9813,"tokens_out":2474,"duration_ms":28002,"significance":"If the reported result is robust, the paper would make a modest empirical contribution: it demonstrates that a small transformer operating on pooled multi-scale CNN features can improve correlation on a standard FBP benchmark, and the ablation study clearly attributes gains to the proposed components. The architecture is simple, clearly described, and reproducible in principle. The authors also explicitly admit the single-fold limitation, which is a positive sign of scientific transparency. However, the headline 'new state-of-the-art' claim is not currently supported by the evidence: the comparison table stops at R3CNN (2019) and omits recent transformer-based FBP methods, the margin over R3CNN is small relative to likely run-to-run variance, and the abstract's claim of outperforming on 'both correlation and error metrics' is contradicted by the paper's own Table 2, where R3CNN has a lower MAE.","major_comments":[{"comment":"The central state-of-the-art claim is supported only by comparisons against a small, largely pre-2020 baseline set. The reference list includes several recent transformer- or attention-based FBP methods, e.g., [13] (Global Context Vision Transformer), [25] (MobileViT), [26] (ViT), and [30] (Mamba-CNN), yet none appear in Table 2. Without protocol-matched results for these methods on the same fold, the statement that SIT 'establishes a new state-of-the-art' is not substantiated. Please either include these baselines under the same evaluation protocol or explicitly justify their exclusion.","section":"Section 4.2, Table 2"},{"comment":"The results are reported on a single fold (fold 3) with no standard deviation, no multiple seeds, and no statistical significance test. The paper's own limitations section admits that averaging across all five folds would be more robust. The advantage over R3CNN is PC +0.0045 and RMSE -0.004, which is within the range one would expect from initialization and training stochasticity. To support the stated ranking, the authors should report mean±std over the five folds (and preferably over multiple runs), and/or provide per-fold results demonstrating that the advantage is consistent.","section":"Section 5.1 and Table 2"},{"comment":"The manuscript repeatedly claims that SIT outperforms prior methods 'in both correlation and error metrics,' but Table 2 shows R3CNN achieves a lower MAE (0.2120 vs. 0.2180). The text in Section 4.2 acknowledges this trade-off, but the abstract and introduction do not. This inconsistency should be corrected, and the authors should state explicitly that SIT improves PC and RMSE at the cost of a slightly higher MAE, or provide evidence that the MAE difference is not material.","section":"Abstract, Section 1 (contribution 4), and Section 4.2"}],"minor_comments":[{"comment":"The spatial dimensions of the 3×3 and 5×5 convolutions are stated as 7×7, but the padding and stride are not specified. For reproducibility, please state these hyperparameters (e.g., same padding, stride 1).","section":"Section 3.2.1, Eqs. (3)–(5)"},{"comment":"The table caption does not specify the fold/split for each baseline. Since fold 3 is used for SIT, please verify and report that all cited baseline numbers correspond to the same evaluation protocol, or note the source of each baseline value.","section":"Table 2"},{"comment":"The MAE and RMSE entries are missing. If these values are unavailable in the original paper, please state that explicitly; otherwise, the dash is ambiguous.","section":"Table 2, 'CNN + LDL' row"},{"comment":"Several methods cited in Related Work ([13], [25], [26], [30]) are closely related to the proposed approach; adding them to the comparison table (or explaining their omission) would also help align the narrative with the reference list.","section":"References"},{"comment":"There are minor typographical issues: the running header 'APREPRINT', the title/header spacing in 'CNN-TRANSFORMERMODEL', and 'de factostandard' should be fixed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the architecture is reasonable, but the empirical support for the central SOTA claim needs substantial strengthening. The reference list contains several recent works by the same author that are directly relevant to the comparison but are omitted from Table 2; this is worth checking for selective citation. The single-fold evaluation is acknowledged by the authors, which is commendable, but it is a load-bearing limitation for the main claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read arXiv:2509.05078. The architecture is a clean hybrid: MobileNetV2 backbone, three parallel convolutions (1x1, 3x3, 5x5), GAP+GMP to produce three 128-d vectors, then a two-block transformer over that sequence. Making scale a sequence dimension and letting attention mix them is a small but legitimate twist on existing hybrid designs. The ablation study is the real asset: it isolates multi-scale features, the transformer, and dual pooling, and each component shows a consistent gain. The writing is honest too—Section 5.1 explicitly admits single-fold evaluation, which is more than many papers do.\n\nThe soft spots are where the paper makes its headline claim. 'New state-of-the-art' rests on Table 2, and that table stops in 2019. It omits the author's own recent transformer-based FBP papers (refs 13 and 30) and any other modern hybrid. The margin over R3CNN is 0.0045 in PC and 0.004 in RMSE—well within run-to-run variance for this dataset. There's no code, no seeds, and no multi-fold average, so I can't tell whether the ordering would survive a second run. The abstract says SIT outperforms 'in both correlation and error metrics,' but Table 2 shows R3CNN's MAE (0.2120) is lower than SIT's (0.2180). That's a factual contradiction in the text.\n\nSo the architecture is plausible and the ablation is useful, but the SOTA claim is not currently supported. The fix is straightforward: report 5-fold averages with error bars, publish code, and compare against the transformer-based methods that are out there.\n\nThis paper deserves a serious referee because the core idea is testable and the ablation gives something to build on. But I'd push for major revision before acceptance—the empirical evidence needs to match the claim. I'd bring it to a reading group only if people are thinking about scale-as-sequence designs; otherwise it's a one-hour skip.","headline":"A clean, honestly-written hybrid architecture whose SOTA claim is under-supported by single-fold results and an incomplete table.","tokens_in":10222,"tokens_out":2542,"would_cite":false,"duration_ms":26823,"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":"The Scale-Interaction Transformer claims a new state-of-the-art for facial beauty prediction, reaching a Pearson correlation of 0.9187 on SCUT-FBP5500 by explicitly modeling interactions between multi-scale facial features.","keywords":["facial beauty prediction","hybrid CNN-transformer","scale-interaction transformer","multi-scale feature extraction","self-attention","SCUT-FBP5500","image regression","computational aesthetics"],"falsifier":"Run SIT and R3CNN on the same folds of SCUT-FBP5500 under identical preprocessing and report the per-fold mean and variance of Pearson correlation; the central claim fails if the average PC over five folds is not above 0.9142, or if the reported fold-3 gap does not reproduce.","tokens_in":9354,"feed_emoji":"😊","tokens_out":7486,"duration_ms":69450,"temperature":0.7,"pith_summary":"The paper proposes the Scale-Interaction Transformer (SIT), a hybrid CNN-transformer built for facial beauty prediction. Its claim is that the missing ingredient in earlier models is not just multi-scale feature extraction, but explicit modeling of how features at different scales interact. On the SCUT-FBP5500 benchmark, SIT reports a Pearson correlation of 0.9187 and an RMSE of 0.2760, which would be a new state of the art, slightly ahead of the region-attentive R3CNN baseline. The paper argues that treating scale-specific features as a short sequence lets self-attention learn the local-to-global relationships that human beauty judgment depends on.","feed_headline":"Scale-interaction transformer scores 0.9187 on facial beauty benchmark","feed_subtitle":"Multi-scale face features become attention tokens, letting self-attention weigh local and global cues together.","key_machinery":"The key object is the Scale-Interaction Transformer (SIT), which converts multi-scale convolutional feature maps into a 3-token sequence and applies transformer self-attention to model their relationships. The three tokens come from parallel 1x1, 3x3, and 5x5 convolutions applied to the MobileNetV2 backbone output, each pooled by global average pooling concatenated with global max pooling. The two-block, four-head transformer encoder then computes attention over these scale tokens, and the sequence is average-pooled before a dense regression head. The mechanism's work is to replace feature fusion (concatenation) with feature interaction (attention), so the final prediction is conditioned on","core_discovery":"The central discovery the paper tries to establish is that a transformer can serve as a scale-interaction module rather than a spatial-interaction module for facial beauty regression. A pre-trained MobileNetV2 backbone produces a 7x7x1280 feature map; three parallel convolutional branches (1x1, 3x3, 5x5) turn it into three scale-specific maps; global average and max pooling compress each map into a 128-dimensional vector; and a two-block transformer encoder with four attention heads learns the dependencies among these three scale tokens. The ablation shows each component adds measurable gain: adding multi-scale features alone raises Pearson correlation from 0.8995 (backbone baseline) to 0.90","pith_inferences":["The same scale-token design could be adapted to apparent age estimation, where global structure and local skin texture interact; this is an extension, not a claim tested in the paper.","If the self-attention weights over the three scales were visualized per image, the model could reveal whether beauty judgments lean on global structure or local texture depending on the face; the paper lists explainability as future work.","Because the paper reports a single fold, averaging all five folds of SCUT-FBP5500 is the natural confirmation step; the stated SOTA margin of 0.0045 in Pearson correlation may shrink or vanish under fold averaging.","The 3-token sequence is far shorter than patch-based ViT sequences, so the architecture offers a computationally cheap way to add global relational reasoning to a CNN for regression, if the single-fold result holds."],"forward_implications":["If SIT's result holds, explicit scale interaction via self-attention is a stronger aggregation strategy than concatenating multi-scale features for facial beauty prediction.","The ablation indicates the transformer block is the largest single contributor, implying attention over scale tokens may help other holistic regression tasks, not just facial beauty.","The gain from adding global max pooling alongside average pooling suggests the dominant and average response of each scale carry non-overlapping signal.","The reported numbers (PC 0.9187, RMSE 0.2760) on fold 3 of SCUT-FBP5500 become the benchmark point that future facial beauty prediction methods would need to beat."],"supporting_citations":[{"why":"Supplies the SCUT-FBP5500 benchmark, the 5-fold protocol, and the human-annotated beauty scores used for all experiments.","marker":"[27]"},{"why":"R3CNN is the previous state-of-the-art baseline whose Pearson correlation, MAE, and RMSE the SIT directly claims to beat.","marker":"[19]"},{"why":"Supplies the transformer encoder and self-attention mechanism that SIT adapts from spatial patch tokens to scale tokens.","marker":"[12]"},{"why":"Provides the parallel multi-scale convolution idea behind the 1x1, 3x3, and 5x5 branches.","marker":"[21]"},{"why":"Defines the alternative feature-pyramid approach to multi-scale fusion that the paper contrasts with interaction modeling.","marker":"[22]"},{"why":"A representative attention-based facial beauty prediction method included in the main comparison table.","marker":"[16]"}],"fun_headline_variants":["Transformer links facial scales for beauty score 0.9187","CNN+Transformer hits 0.9187 on beauty prediction","Scale-interaction transformer tops beauty benchmark","Hybrid model reads face scales for beauty prediction","Multi-scale face tokens boost beauty score to 0.9187"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The new-state-of-the-art claim rests on the assumption that the prior results in Table 2 were obtained under exactly the same fold, image preprocessing, and metric conventions; the paper itself states that only a single fold is evaluated, so a small protocol mismatch could erase the 0.0045 advantage.","fun_headline_variants_meta":{"raw":{"variants":["Transformer links facial scales for beauty score 0.9187","CNN+Transformer hits 0.9187 on beauty prediction","Scale-interaction transformer tops beauty benchmark","Hybrid model reads face scales for beauty prediction","Multi-scale face tokens boost beauty score to 0.9187"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1054,"prompt_tokens":771,"completion_tokens":283,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":204}},"tokens_in":515,"tokens_out":283,"duration_ms":2902,"temperature":1.0,"reasoning_tokens":204,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:36:22.909672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SIT and R3CNN on the same folds of SCUT-FBP5500 under identical preprocessing and report the per-fold mean and variance of Pearson correlation; the central claim fails if the average PC over five folds is not above 0.9142, or if the reported fold-3 gap does not reproduce.","supporting_citations":[{"cited_title":"Liang, L","cited_arxiv_id":null,"evidence_quote":"Supplies the SCUT-FBP5500 benchmark, the 5-fold protocol, and the human-annotated beauty scores used for all experiments."},{"cited_title":"Regression Guided by Relative Ranking Using Convolutional Neural Network (R3CNN) for Facial Beauty Prediction","cited_arxiv_id":null,"evidence_quote":"R3CNN is the previous state-of-the-art baseline whose Pearson correlation, MAE, and RMSE the SIT directly claims to beat."},{"cited_title":"Rethinking the inception architecture for computer vision","cited_arxiv_id":null,"evidence_quote":"Provides the parallel multi-scale convolution idea behind the 1x1, 3x3, and 5x5 branches."},{"cited_title":"Feature pyramid networks for object detection","cited_arxiv_id":null,"evidence_quote":"Defines the alternative feature-pyramid approach to multi-scale fusion that the paper contrasts with interaction modeling."},{"cited_title":"Cao, K Choi, H Jung et al., Deep learning for facial beauty prediction","cited_arxiv_id":null,"evidence_quote":"A representative attention-based facial beauty prediction method included in the main comparison table."}],"review_version":1}