{"id":"9bda9009-bf9a-46a2-9538-25343e05cc44","arxiv_id":"2411.10863","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Diffusion-generated synthetic facial images are reported to raise FER2013 accuracy to 96.47% and RAF-DB accuracy to 99.23% for ResEmoteNet.","lead":"This paper reports that training ResEmoteNet on FER2013 and RAF-DB supplemented with many Stable Diffusion generated face images yields accuracies of 96.47% and 99.23%, far above prior published results. The significance, if confirmed, is that synthetic data can fix emotion-class imbalance, but the paper's evaluation protocol is too vaguely described to trust the numbers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The absence of a specified validation/test partition and FER2013 public-vs-private split undermines every reported accuracy gain; the unaugmented baseline already exceeds the listed SOTA, indicating a protocol mismatch.","rationale":"The reader's weakest assumption is exactly the condition on which the central claim depends. The paper reports near-perfect accuracy values without specifying which FER2013 test split is used or how the early-stopping validation set is carved out. The internal evidence is strong: the unaugmented baseline already beats all SOTA numbers in Table IV, which is difficult to explain unless the protocol differs from the comparison methods. This is not a disagreement with consensus; it is an unverified empirical claim where the evaluation protocol has a hole. The diffusion-generation pipeline itself is described in reasonable detail, and the authors do link the ResEmoteNet code, but neither the synthetic generation prompts nor the exact data splits are released, so the reported numbers cannot be independently checked. Given that the headline contribution is the accuracy gain itself, this missing protocol information is fatal to the claim as stated. I therefore see no reason to change the reader's REJECT verdict.","tokens_in":8380,"tokens_out":4401,"duration_ms":47137,"concrete_test":"Obtain the exact configuration from the ResEmoteNet repository or the authors: identify the FER2013 test split used for Table III (public or private), the exact validation split used for early stopping, and the number of validation samples for RAF-DB. Then re-run the baseline and Augmentation 4 under the official FER2013 split with disjoint validation (e.g., public test) and final test (private test), reporting both public and private accuracies. If the validation partition overlaps the reported test set by any sample, recompute all Table III/IV numbers under a fully disjoint split; the central claim stands only if the augmented-minus-baseline gap remains large and the SOTA comparison uses the same split.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that 15,000 synthetic images per class raises accuracy to 96.47% on FER2013 and 99.23% on RAF-DB—rests entirely on the evaluation protocol. Section IV (Training Configuration) reports early stopping based on validation loss with patience 5, but never states how the validation set is formed from the FER2013 training split (28,709 images) or the RAF-DB training split (12,271 images), nor whether FER2013 results use the public test or private test partition. Section III's statement that 'the validation and testing sets remained unchanged' does not resolve this because no initial validation partitioning is described. FER2013 has distinct public and private test sets of 3,589 images each, and their accuracies differ materially; early stopping on the same partition later reported as test accuracy is a validation leak. The strongest internal red flag is Table IV: the unaugmented ResEmoteNet baseline (79.79%) already surpasses every listed SOTA method (highest 77.85% Multi-Branch ViT, 76.82% Ensemble ResMaskingNet). Either the baseline is evaluated on a different, easier partition than the comparison methods, or the comparison table is not on the same protocol; either way the claimed improvements and SOTA superiority are unverifiable. Missing seeds, error bars, and generated-data release further limit verification, but the split/protocol issue is the load-bearing one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-augmentation method for facial emotion recognition (FER) that uses Stable Diffusion 2 and Stable Diffusion 3 Medium to generate synthetic per-class training images, which are added to the FER2013 and RAF-DB training sets. The authors then train ResEmoteNet, their own previously proposed architecture, on the augmented data and report accuracies of 96.47% on FER2013 and 99.23% on RAF-DB, compared with 79.79% and 94.76% without augmentation. They compare these results with a set of prior methods in Table IV and conclude that diffusion-based synthetic augmentation substantially improves FER accuracy.","tokens_in":8641,"tokens_out":9453,"duration_ms":85638,"significance":"If the reported numbers are correct, the paper would demonstrate a striking benefit of text-to-image diffusion models for FER, with gains of roughly 17% and 5% on two standard benchmarks. The confusion matrices in Figure 4 are internally consistent, and the idea of balancing classes with synthetic data is well motivated. However, the paper provides no code for the augmentation pipeline, no seeds or error bars, no verification of synthetic label quality, and no control that separates the augmentation effect from the choice of ResEmoteNet. The significance of the empirical claim is therefore currently limited by reproducibility and evaluation concerns.","major_comments":[{"comment":"The evaluation protocol is not defined. Section IV says the model is trained for up to 80 epochs with early stopping on validation loss (patience 5), but never states how the validation set is formed from the 28,709 FER2013 training images or the 12,271 RAF-DB training images. Section III's sentence 'The validation and testing sets remained unchanged' is uninterpretable without an initial partition. The paper also does not state whether the FER2013 test accuracy is on the public or the private test split. As a result, the central claim of a 16.68% absolute improvement on FER2013 is not verifiable. The internal inconsistency in Table IV, where the unaugmented ResEmoteNet baseline (79.79%) already exceeds all listed SOTA methods (max 77.85%), strengthens the concern that the baseline and the comparison methods use different evaluation protocols.","section":"Section IV (Training Configuration) and Section III (Diffusion Based Data Augmentation)"},{"comment":"All results in Table III and Figure 4 are single-run point estimates. Because both the diffusion sampling and the network initialization are stochastic, the reported accuracies are meaningless without a measure of variance. The authors should provide the mean and standard deviation over at least three independent runs, and state the random seeds used for synthetic-data generation and training. This is necessary to determine whether the differences between Aug. 1 and Aug. 4 (e.g., 84.81% versus 96.47% on FER2013) are significant or within run-to-run noise.","section":"Section V (Results and Discussion)"},{"comment":"The paper provides no verification that the synthetic images actually express the intended emotion. The prompts listed in Table I are generic and the text does not describe how many prompts were used, how many images were generated per prompt, or how the 15,000-images-per-class target was reached without excessive duplication. Without a manual or automated check of a sample of generated images, and without an analysis of the synthetic distribution's similarity to the real distribution, the accuracy gains in Table III could be attributed to mislabeled synthetic examples or to overfitting to synthetic artifacts rather than to corrected class balance.","section":"Section III (Diffusion Based Data Augmentation) and Figure 3"},{"comment":"The experimental design cannot separate the effect of diffusion-based augmentation from the choice of ResEmoteNet as the backbone, because all experiments use ResEmoteNet only. Since ResEmoteNet is the authors' own prior architecture, a control experiment with a standard FER network (e.g., ResNet-18) trained on the same augmented datasets is necessary to support the conclusion that synthetic data augmentation is what drives the improvement. Without such a control, the statement in Section VI that synthetic data is 'instrumental' is not established.","section":"Section V.A (Comparison with previous study) and Table IV"}],"minor_comments":[{"comment":"There are figure reference errors: Section II.A says 'Fig. II' but the model diagram is Figure 1, and Section III says 'as shown in Fig. IV' but the synthetic images appear in Figure 3.","section":"Section II.A and Section III"},{"comment":"Equation (1) introduces ws as the attention weights from the SE Block but does not specify its dimensions or the type of multiplication (element-wise or matrix); please clarify the notation.","section":"Section II.A, Eq. (1)"},{"comment":"There is a grammatical error: 'These findings shows' should be 'These findings show.'","section":"Abstract"},{"comment":"The paper states that synthetic images are resized to 64x64 'to align with the input requirements of the model,' but it is not stated that the real images are also resized to this resolution; please specify the input resolution for all training and test images.","section":"Section IV (Data Processing)"},{"comment":"The caption 'Aug. refers to augmentation' is uninformative; the table should state explicitly that the entries are target per-class sample counts after augmentation and define what 'Both' means in the Aug. 2-4 columns.","section":"Table II"},{"comment":"No hyperparameters are given for the diffusion models used in data generation (e.g., number of denoising steps, guidance scale, or random seeds), so the synthetic data generation is not reproducible.","section":"Section IV (Training Configuration)"}],"recommendation":"reject","confidential_remarks":"The paper is heavily dependent on the authors' own ResEmoteNet architecture, and the evaluation details are insufficient. The unexplained gap between the unaugmented baseline and all listed SOTA methods in Table IV is a particular concern; if the authors are using a non-standard split, the central comparison would be invalid. I recommend rejection unless the authors can provide a clear protocol description and additional control experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does a straightforward thing—generate synthetic FER images with Stable Diffusion, add them to the training set, train ResEmoteNet—and reports very large accuracy jumps (79.79 to 96.47 on FER2013, 94.76 to 99.23 on RAF-DB). Those numbers would matter if they hold. But the evaluation setup is missing enough detail that I can't tell whether they hold.\n\nWhat's new: applying diffusion-based augmentation to FER with a specific scaling recipe (10k–15k synthetic images per class) and reporting the resulting numbers. The underlying idea is established (they cite [8], [9]), so this is an extension, not a breakthrough. What the paper does well: the pipeline is clearly described, class-wise accuracies and confusion matrices are given for the main results, and the model code is public. That's more than many short papers provide.\n\nThe soft spots are real and load-bearing. First, the validation set is never defined. They say they use early stopping on validation loss with patience 5, but never state how the validation split is created from the training data. FER2013 has public and private test sets; the paper doesn't say which one is used. If the validation set overlaps with the test set, or if the reported test accuracy is from a different partition than the comparisons, every gain in Table III is inflated. Second, Table IV is internally suspicious: the unaugmented ResEmoteNet baseline (79.79% on FER2013) already beats every listed SOTA method (highest 77.85%). That strongly suggests a protocol mismatch—either the baseline is evaluated on an easier partition or the SOTA numbers are from different splits. Either way, the comparison table doesn't establish superiority.\n\nMissing seeds, error bars, and synthetic-image verification are also problems, but they're secondary. If the split issue is resolved and the numbers hold, the paper is a solid data-centric augmentation result. As written, it's unverifiable.\n\nThis paper is for a reader interested in whether diffusion-based augmentation can fix class imbalance in FER. It deserves a serious referee because the question is worth answering, but the referee should require a precise split description, public/private test designation, error bars over multiple runs, and a comparison to simpler augmentation baselines. I'd send it to review with those demands, not desk-reject it.","headline":"The paper's diffusion-augmentation recipe is a reasonable thing to try and the gains are big, but the evaluation protocol is so under-specified that the headline numbers can't be taken at face value.","tokens_in":9187,"tokens_out":2419,"would_cite":false,"duration_ms":24883,"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":"Adding synthetic faces generated by Stable Diffusion to the training set raises facial emotion recognition accuracy from 79.79% to 96.47% on FER2013 and from 94.76% to 99.23% on RAF-DB.","keywords":["facial emotion recognition","synthetic data augmentation","diffusion models","Stable Diffusion","class imbalance","ResEmoteNet","FER2013","RAF-DB"],"falsifier":"Retrain ResEmoteNet under the same augmentation protocol but with the validation set taken only from the original training split and test accuracy computed on the official FER2013 private test set and the RAF-DB test set; if the gap over the unaugmented baseline shrinks materially below 16.68 and 4.47 percentage points, the central claim would need revision.","tokens_in":8154,"feed_emoji":"😀","tokens_out":9640,"duration_ms":88157,"temperature":0.7,"pith_summary":"The paper sets out to show that the class imbalance in facial emotion recognition can be corrected by adding synthetic face images produced by diffusion models to the training set. Using ResEmoteNet, a convolutional network built from squeeze-and-excitation blocks and residual connections, the authors generate images with Stable Diffusion 2 and Stable Diffusion 3 Medium from text prompts and balance every emotion class to 15,000 samples. They report that this raises accuracy from 79.79% to 96.47% on FER2013 and from 94.76% to 99.23% on RAF-DB, an absolute gain of 16.68 and 4.47 percentage points respectively. The significance is that unbalanced emotion datasets could be repaired synthetically rather than by collecting more labels, and that the reported numbers exceed the comparison methods in the paper's table.","feed_headline":"Synthetic faces lift emotion recognition to 96.47 percent","feed_subtitle":"Balancing FER2013 with 15,000 generated images per class adds 16.68 points; RAF-DB reaches 99.23 percent.","key_machinery":"The load-bearing mechanism is ResEmoteNet, a convolutional network that combines squeeze-and-excitation blocks with residual skip connections, trained on diffusion-augmented data. The squeeze-and-excitation blocks recalibrate channel-wise feature responses, the residual blocks preserve the input signal to ease training, and the augmentation step, using Stable Diffusion 2 and Stable Diffusion 3 Medium, generates new facial images from text prompts. The operation that carries the argument is the systematic balancing of class sizes: FER2013 and RAF-DB are augmented in stages up to 15,000 images per class, and accuracy rises monotonically with each stage.","core_discovery":"The paper's claim is that diffusion-based synthetic augmentation, not a new architecture, is what drives the accuracy jump. The pipeline generates synthetic facial expressions from text prompts, resizes the generated images to the model input size, and merges them into the existing training distributions of FER2013 and RAF-DB so each of the seven emotion classes has up to 15,000 samples. On FER2013 the reported test accuracy rises from 79.79% to 96.47%, with the Fear class improving from 70% to 95% and Happy reaching 100%; on RAF-DB accuracy rises from 94.76% to 99.23%, with four classes at 100%. The authors interpret these results as evidence that balancing the class distribution with synthetic data makes the model generalize better across all emotions.","pith_inferences":["A direct test of the claim would be to repeat the augmentation on a different backbone, such as a standard ResNet, to see whether the accuracy jump is specific to ResEmoteNet or transfers to other architectures.","Because several classes reach 100% accuracy after augmentation, an independent check on a truly held-out test set would reveal whether this reflects genuine generalization or a distributional overlap between synthetic training images and test images.","Comparing 48x48 resizing against the 64x64 resizing used here would clarify whether the improvement comes from class balancing or from a shift in input resolution."],"forward_implications":["On FER2013, augmenting each emotion class to 15,000 samples yields 96.47% accuracy, with Happy at 100% and Fear up from 70% to 95%.","On RAF-DB, the same augmentation reaches 99.23% accuracy, with Happy, Neutral, Sad, and Surprise all at 100%.","The augmented ResEmoteNet outperforms the comparison methods listed in the paper's Table IV on both datasets.","The reported gains are largest for the classes that were least represented originally, indicating that balancing alone accounts for much of the improvement.","Training remains practical: roughly 6.5 hours for FER2013 and 5 hours for RAF-DB on a Tesla P100, with inference under 5 ms per image."],"supporting_citations":[{"why":"Defines ResEmoteNet, the architecture whose unaugmented and augmented accuracies form the paper's result.","marker":"[11]"},{"why":"Stable Diffusion 2 generates part of the synthetic facial data used to balance the training sets.","marker":"[12]"},{"why":"Stable Diffusion 3 Medium generates the other part of the synthetic facial data.","marker":"[13]"},{"why":"Supplies the FER2013 benchmark, its class distribution, and the test data for reported accuracy.","marker":"[19]"},{"why":"Supplies the RAF-DB benchmark and its single-label training and test subsets.","marker":"[20]"},{"why":"Provides a prior FER2013 result that the paper's augmented model claims to surpass.","marker":"[1]"},{"why":"Provides a prior RAF-DB result that the paper's augmented model claims to surpass.","marker":"[2]"}],"fun_headline_variants":["Fake faces, real gains: emotion AI hits 96.47%","Diffusion-generated faces boost emotion recognition to 96.47%","Balancing emotions with synthetic faces lifts FER to 96.47%","Synthetic emotion data sharpens AI to 99.23% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume that the validation set used for early stopping is drawn only from the training portion and that the test accuracy is measured on the same FER2013 split used by the comparison methods; the paper does not specify either choice.","fun_headline_variants_meta":{"raw":{"variants":["Fake faces, real gains: emotion AI hits 96.47%","Diffusion-generated faces boost emotion recognition to 96.47%","Balancing emotions with synthetic faces lifts FER to 96.47%","Synthetic emotion data sharpens AI to 99.23% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000965,"raw_usage":{"total_tokens":4109,"prompt_tokens":952,"completion_tokens":3157,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":3077}},"tokens_in":568,"tokens_out":3157,"duration_ms":25978,"temperature":1.0,"reasoning_tokens":3077,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:12:50.639993+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain ResEmoteNet under the same augmentation protocol but with the validation set taken only from the original training split and test accuracy computed on the official FER2013 private test set and the RAF-DB test set; if the gap over the unaugmented baseline shrinks materially below 16.68 and 4.47 percentage points, the central claim would need revision.","supporting_citations":[{"cited_title":"Scaling rectified flow transformers for high-resolution image synthesis,","cited_arxiv_id":null,"evidence_quote":"Stable Diffusion 3 Medium generates the other part of the synthetic facial data."},{"cited_title":"Challenges in representation learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the FER2013 benchmark, its class distribution, and the test data for reported accuracy."},{"cited_title":"Reliable crowdsourcing and deep locality- preserving learning for expression recognition in the wild,","cited_arxiv_id":null,"evidence_quote":"Supplies the RAF-DB benchmark and its single-label training and test subsets."},{"cited_title":"Facial expression recognition using residual masking network,","cited_arxiv_id":null,"evidence_quote":"Provides a prior FER2013 result that the paper's augmented model claims to surpass."},{"cited_title":"A dual-direction attention mixed feature network for facial expression recognition,","cited_arxiv_id":null,"evidence_quote":"Provides a prior RAF-DB result that the paper's augmented model claims to surpass."}],"review_version":1}