{"id":"e04b1524-3b84-4789-b66a-e386453641c9","arxiv_id":"2507.20363","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"On FBP5500, Diff-FBP reports a Pearson correlation of 0.9220 and MAE of 0.2110 using a frozen Diffusion Transformer feature extractor with generative pre-training.","lead":"This paper introduces Diff-FBP, which pre-trains a Diffusion Transformer on unlabeled face images and uses its frozen features to predict facial beauty scores, reporting a new state-of-the-art on FBP5500. The key claim is that generative pre-training on faces transfers better to subjective aesthetic tasks than standard ImageNet pre-training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation compares ViT-Base and DiT, so generative pre-training is not isolated as the cause of the +0.021 PCC gain.","rationale":"The paper's strongest claim is that Diff-FBP achieves SOTA and that generative pre-training is the key contributor. The ablation in Table 2 is the only evidence for the latter. But the comparison between ViT-Base (ImageNet) and Diff-FBP (generative DiT) changes three variables at once, so the +0.021 PCC gain cannot be attributed to pre-training. This is the load-bearing weakness because the paper's contribution is the paradigm shift, not the architecture. The proposed test—pre-training the same ViT backbone with both a diffusion objective on FFHQ and classification on ImageNet—directly isolates the pre-training task. If the diffusion-pretrained ViT fails to beat the ImageNet-pretrained ViT, the central claim collapses. We agree with the reader's weakest assumption and verdict; the paper should not be accepted without a controlled comparison and code release.","tokens_in":62,"tokens_out":6731,"duration_ms":79250,"concrete_test":"Perform a controlled ablation with architecture held constant. Use the same ViT-Base/16 backbone and pre-train it under three conditions: (a) random initialization, (b) ImageNet-1k classification, and (c) DDPM denoising on FFHQ. Fine-tune the identical two-layer MLP regression head on FBP5500's five folds with the same optimizer, learning rate, and epochs as described in §4.1. If the denoising-pretrained ViT (c) does not outperform the ImageNet-pretrained ViT (b) by a comparable margin (~+0.02 PCC), the paper's claimed causality is refuted. Report the mean and standard deviation of PCC/MAE over the five folds for each condition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central attribution claim relies on Table 2, but rows (2) and (4) differ in architecture family (ViT-Base vs DiT), parameter count, and pre-training objective. The paper states in §4.4 that comparing (2) and (4) 'isolates the effect of the pre-training task,' which is not valid because the backbone is not held constant. The +0.021 PCC gain could come from the DiT's adaLN-Zero blocks, its larger capacity, or the fine-tuning protocol rather than from generative pre-training. Since the paper's novelty is the generative pre-training framework, this confound directly undermines the central claim. Other issues, such as the PCC discrepancy between the abstract (0.932) and the body (0.9220) and the reference to a '[CLS]' token that does not exist in the described DiT architecture, further weaken confidence but are secondary.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Diff-FBP, a two-stage framework for facial beauty prediction. In the first stage, a Diffusion Transformer is pre-trained on FFHQ with a DDPM denoising objective; in the second stage, the DiT encoder is frozen and only a lightweight MLP regression head is trained on FBP5500 under the standard 5-fold cross-validation protocol. The authors report a Pearon Correlation Coefficient of 0.9220 and a Mean Absolute Error of 0.2110, claim a new state-of-the-art on FBP5500, and attribute the improvement to generative pre-training based on an ablation that compares an ImageNet-pretrained ViT-Base with the generatively pre-trained DiT.","tokens_in":9206,"tokens_out":5680,"duration_ms":58821,"significance":"If properly established, the proposed direction would be a meaningful contribution: replacing generic ImageNet pre-training with domain-specific generative pre-training for subjective visual tasks is a plausible and potentially reusable idea, and the two-stage framework is simple to understand. The manuscript uses a standard benchmark, includes a candid limitations section on computational cost and dataset bias, and makes a falsifiable empirical claim. However, the current evidence does not isolate the effect of generative pre-training, the headline result is internally inconsistent, and the comparison set is dated. With a controlled ablation and more complete reporting, the work could become a useful contribution to facial beauty prediction and representation learning.","major_comments":[{"comment":"The claim that comparing rows (2) and (4) 'isolates the effect of the pre-training task' is not valid, because the two rows differ simultaneously in architecture family (ViT-Base versus DiT), parameter count, pre-training dataset (ImageNet versus FFHQ), and fine-tuning protocol; the +0.021 PCC gain could therefore come from the DiT's adaLN-Zero blocks, larger capacity, or different fine-tuning procedure rather than from generative pre-training. I request a controlled comparison that holds the backbone fixed (for example, an ImageNet-pretrained DiT versus the generatively pre-trained DiT, or a ViT-Base with generative pre-training versus ImageNet pre-training), with multiple seeds and error bars.","section":"Section 4.4, Table 2"},{"comment":"The headline result is reported inconsistently as PCC 0.932 in the abstract and PCC 0.9220 in Table 1, Section 4.3, and the conclusion; this discrepancy must be resolved and the correct value used consistently throughout the manuscript.","section":"Abstract, Section 4.3, Conclusion"},{"comment":"The comparison set appears incomplete for a 'new state-of-the-art' claim, as the table lists methods up to R3CNN (2019) and omits recent FBP works and generic pre-trained transformer baselines evaluated on FBP5500 in the intervening years; please expand the comparison to current methods or temper the state-of-the-art claim accordingly.","section":"Section 4.3, Table 1"},{"comment":"The experimental section omits details needed to assess and reproduce the results, including the DiT configuration (depth, width, patch size), number of pre-training epochs, batch size, input resolution, regression-head hyperparameters, per-fold results, and any variance or significance estimate; without error bars, the reported PCC of 0.9220 cannot be distinguished from the 0.901 ablation baseline with statistical confidence.","section":"Sections 4.1-4.4"}],"minor_comments":[{"comment":"The regression head is described as taking the global '[CLS]' token embedding, but the DiT architecture described in Section 3.2 and Algorithm 1 does not define a [CLS] token; please specify how the sequence of patch tokens is pooled for regression.","section":"Section 3.3"},{"comment":"Equation (2) and nearby text contain formatting artifacts, including missing spaces and run-together symbols; please proofread the manuscript.","section":"Equation (2) and surrounding text"},{"comment":"The 'trained from scratch' comparison also changes the architecture between rows (1) and (3), so the statement that these rows demonstrate the effect of pre-training should be reworded or supported with architecture-matched scratch baselines.","section":"Table 2, rows (1) and (3)"},{"comment":"The claims about applicability to medical imaging and other subjective tasks are speculative and unsupported by experiments; please move them to a clearly labeled hypothesis or future-work subsection.","section":"Section 5.2"},{"comment":"References [14], [18], [20], [21], and [26] are the authors' own works; please verify that each is necessary and avoid excessive self-citation.","section":"References"},{"comment":"The statement that fine-tuning 'took only a few minutes per fold' is not a reproducible timing description; please report the exact hardware and training time.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central claim is not supported as written because the key ablation is confounded. I considered rejection, but the flaw is fixable with additional controlled experiments, so I recommend major revision rather than rejection. I also noted the high rate of self-citations; it did not affect my technical assessment, but the authors should trim unnecessary self-references."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper applies a known recipe — diffusion-model pre-training, frozen encoder, lightweight head — to facial beauty prediction. That is a legitimate new application, and the 'to our knowledge' claim is probably right. The writing is clear, and the limitations section is honest about compute cost and dataset bias. I also appreciate that the authors don't oversell the broader reach beyond a few speculative paragraphs.\n\nThe problem is the evidence. The headline PCC is 0.932 in the abstract and 0.9220 in the body; that is not a typo-level inconsistency. Table 2, which is supposed to 'isolate the effect of the pre-training task,' compares ViT-Base (ImageNet) with DiT (generative) — different architecture family, parameter count, and fine-tuning protocol. So the +0.021 gain cannot be pinned on generative pre-training. The §4.4 statement that this comparison uses the 'same ViT-style architecture' is wrong. There are also no error bars, no variance across the 5 folds, and the baseline list omits several recent methods, so the SOTA claim is fragile. The reference to a global '[CLS]' token also sits oddly with a DiT that produces per-patch outputs.\n\nThese flaws are not fatal to the underlying idea — a controlled comparison with a DiT trained on ImageNet, or a linear probe on the same backbone, would settle it. Code and exact settings would help. But as written, the central attribution claim is not supported.\n\nWho is this for? Anyone interested in generative pre-training for subjective tasks. The paper deserves a serious referee only if the authors can run the right controls. I would not send this version out for review; I'd invite a resubmission after a controlled study.\n\nRecommendation: desk reject in present form, with a concrete path to revision.","headline":"A plausible application of DiT pre-training to beauty prediction is undercut by a confounded ablation and an inconsistent headline number.","tokens_in":9742,"tokens_out":3799,"would_cite":false,"duration_ms":40193,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage framework pre-trains a Diffusion Transformer to denoise faces, then freezes its encoder as a feature backbone; the paper reports a state-of-the-art PCC of 0.9220 and MAE of 0.2110 on FBP5500.","keywords":["facial beauty prediction","diffusion transformer","generative pre-training","self-supervised learning","transfer learning","FBP5500","FFHQ","subjective visual assessment"],"falsifier":"Train the same DiT architecture with the same frozen-encoder and regression-head protocol, but initialize it with ImageNet classification pre-training instead of face denoising; if its FBP5500 PCC stays near 0.922 rather than dropping toward 0.901, the generative-pre-training claim fails. Separately, the abstract's 0.932 and the results table's 0.9220 cannot both be the reported PCC, so checking the raw predictions against one of these numbers would settle which result is real.","tokens_in":8869,"feed_emoji":"📈","tokens_out":5601,"duration_ms":58056,"temperature":0.7,"pith_summary":"This paper argues that the dominant pre-training recipe for facial beauty prediction—starting from an ImageNet classification model—is mismatched to the task, and that pre-training a model to generate faces is a better foundation. It introduces Diff-FBP, a two-stage framework that first trains a Diffusion Transformer to denoise unlabeled face images and then freezes that encoder as a feature extractor, training only a lightweight regression head on the FBP5500 benchmark. The paper reports a new state of the art, with Pearson Correlation Coefficient 0.9220 and Mean Absolute Error 0.2110, and claims its ablations show the generative pre-training phase, not the architecture, is responsible. A sympathetic reader would care because the recipe promises a way to build strong models for subjective visual tasks from small labeled datasets by using unlabeled domain images instead of generic classification features.","feed_headline":"Faces, not objects: diffusion pre-training sets beauty-score record","feed_subtitle":"Generative pre-training on faces, then freezing the encoder, beats ImageNet backbones on FBP5500 with PCC 0.922.","key_machinery":"The engine is a Diffusion Transformer (DiT): a Vision-Transformer-style network that takes a noisy image patch sequence and a timestep embedding, and predicts the added noise. Its DiT blocks use adaLN-Zero, adaptive layer normalization in which the timestep conditions scale and shift each layer's activations, with zero-initialized projections so each block starts as an identity. The same encoder, frozen after denoising pre-training, outputs the token features that feed the regression head; the surrounding machinery is the two-stage protocol itself—denoise to learn the face manifold, then freeze and fit only a small head.","core_discovery":"Diff-FBP establishes that reconstructive features—learned by predicting noise in a denoising diffusion process over 70,000 unlabeled faces from FFHQ—transfer better to aesthetic scoring than discriminative features learned by classifying ImageNet objects. With the Diffusion Transformer encoder frozen and only a two-layer MLP regression head trained, the model reaches PCC 0.9220 and MAE 0.2110 on FBP5500 under the standard 5-fold protocol, outperforming the previous best method at 0.9142 PCC. The ablation pairs an ImageNet-pretrained ViT-Base (PCC 0.901) against the generative-pretrained Diff-FBP (PCC 0.922) and attributes the gain to the pre-training objective; the paper's stated conclusion is that learning the data manifold of faces supplies the holistic structural priors that human aesthetic judgment relies on.","pith_inferences":["A cleaner test of the causal claim would compare two models matched in architecture and fine-tuning protocol that differ only in pre-training objective (face denoising vs. ImageNet classification); the current Table 2 varies architecture and protocol at the same time.","The paper reports PCC 0.932 in the abstract and 0.9220 in the results table and conclusion; reconciling that discrepancy matters before comparing against other published numbers.","If the mechanism generalizes, an immediate testable extension is to pre-train a DiT on unlabeled medical images and measure whether downstream pathology classification improves over ImageNet backbones."],"forward_implications":["If the central claim is right, the new FBP5500 state of the art shifts from classification-pretrained CNNs and Transformers to a generative-pretrained frozen encoder, at PCC 0.9220 and MAE 0.2110.","Downstream fine-tuning becomes cheap: only the regression head is trained, taking minutes per fold on a single GPU, so the framework lowers the labeled-data burden for beauty prediction.","The same two-stage recipe is proposed for other subjective visual tasks, including medical image analysis, computational aesthetics, and product design.","The results suggest that features optimized to reconstruct a domain are more semantically aligned with subjective human ratings than features optimized to discriminate object categories."],"supporting_citations":[{"why":"Supplies the DDPM denoising objective used for the generative pre-training phase.","marker":"[16]"},{"why":"Defines the Diffusion Transformer architecture with adaLN-Zero used as the encoder.","marker":"[11]"},{"why":"Provides the FBP5500 benchmark and the 5-fold cross-validation evaluation protocol.","marker":"[9]"},{"why":"Supplies the 70,000 unlabeled face images used in the pre-training phase.","marker":"[19]"},{"why":"ResNet-50 is one of the classification-pretrained baselines compared in Table 1.","marker":"[6]"},{"why":"R3CNN is the previous state-of-the-art method that Diff-FBP is claimed to outperform.","marker":"[31]"},{"why":"ViT-Base is the ImageNet-pretrained architecture used as the ablation control.","marker":"[7]"},{"why":"ImageNet is the generic classification pre-training source that the paper contrasts with generative pre-training.","marker":"[8]"}],"fun_headline_variants":["Diffusion pre-training on faces beats ImageNet for beauty scoring","Frozen diffusion encoder beats ImageNet backbones in beauty prediction","Generative pre-training on faces beats object classification for beauty","Denoising faces beats classifying objects for beauty scores","Diffusion pre-training lifts beauty prediction over ImageNet baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The causal claim rests on the ablation in Table 2, which compares an ImageNet-pretrained ViT-Base with the generative-pretrained Diff-FBP; because the two models differ in architecture, parameter count, and fine-tuning protocol, the +0.021 PCC gain may come from those differences rather than from generative pre-training.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion pre-training on faces beats ImageNet for beauty scoring","Frozen diffusion encoder beats ImageNet backbones in beauty prediction","Generative pre-training on faces beats object classification for beauty","Denoising faces beats classifying objects for beauty scores","Diffusion pre-training lifts beauty prediction over ImageNet baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001261,"raw_usage":{"total_tokens":5193,"prompt_tokens":1000,"completion_tokens":4193,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":4110}},"tokens_in":616,"tokens_out":4193,"duration_ms":36254,"temperature":1.0,"reasoning_tokens":4110,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:33:33.794794+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same DiT architecture with the same frozen-encoder and regression-head protocol, but initialize it with ImageNet classification pre-training instead of face denoising; if its FBP5500 PCC stays near 0.922 rather than dropping toward 0.901, the generative-pre-training claim fails. Separately, the abstract's 0.932 and the results table's 0.9220 cannot both be the reported PCC, so checking the raw predictions against one of these numbers would settle which result is real.","supporting_citations":[{"cited_title":"”Denoising diffusion probabilistic models.” Advances in neural information processing systems 33 (2020): 6840-6851","cited_arxiv_id":null,"evidence_quote":"Supplies the DDPM denoising objective used for the generative pre-training phase."},{"cited_title":"”Scalable diffusion models with transformers.” Proceed- ings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Defines the Diffusion Transformer architecture with adaLN-Zero used as the encoder."},{"cited_title":"Liang, L","cited_arxiv_id":null,"evidence_quote":"Provides the FBP5500 benchmark and the 5-fold cross-validation evaluation protocol."},{"cited_title":"”Diverse Dataset for Eyeglasses Detection: Extending the Flickr- Faces-HQ (FFHQ) Dataset.” Sensors 24.23 (2024): 7697","cited_arxiv_id":null,"evidence_quote":"Supplies the 70,000 unlabeled face images used in the pre-training phase."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ResNet-50 is one of the classification-pretrained baselines compared in Table 1."},{"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 method that Diff-FBP is claimed to outperform."},{"cited_title":"”Imagenet: A large-scale hierarchical image database.” 2009 IEEE con- ference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"ImageNet is the generic classification pre-training source that the paper contrasts with generative pre-training."}],"review_version":1}