{"id":"8c9eb8ae-2341-4597-ac32-7ea052ddaba9","arxiv_id":"2412.07196","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FG-RAT GAN, a RAT GAN enhanced with an auxiliary classifier and contrastive learning over cross-batch memory, reports lower FID than its backbone on two fine-grained datasets.","lead":"This paper adds an auxiliary classifier and a contrastive learning loss to the RAT GAN text-to-image model, aiming to make generated images respect fine-grained class differences. On the CUB bird and Oxford flower benchmarks, the modified model reports lower FID scores than the base RAT GAN, though the comparison is not fully controlled.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 3/4's negative-pair term has the wrong sign: as printed, minimizing it maximizes cross-class cosine similarity, contradicting the stated contrastive objective.","rationale":"The reader's stated weakest assumption is the label-supervision asymmetry in the comparison against label-free baselines. That is a legitimate concern, but the manuscript contains a more direct, checkable inconsistency in the core methodology: the negative-pair term in Eqs. (3) and (4) has the opposite sign from what the prose claims. Since the contrastive loss is one of the two principal contributions, this sign error threatens the internal validity of the central claim far more immediately than the fairness of the baseline comparison. The FID improvements reported in Table 2 could still be real, but only if the implementation silently corrected the sign; without code, the paper as written is not reproducible. The correct remedy is to require the authors to fix the equation, release code, and report multiple seeds. That is exactly the CONDITIONAL posture the reader already took, so the verdict should remain UNCHANGED, but for a different and more precise reason than the one the reader emphasized.","tokens_in":896,"tokens_out":893,"duration_ms":66176,"concrete_test":"Reproduce Eq. (3) on a single negative pair with cos_sim = 0.8 and alpha = 0.5. As printed, the term is -(0.3) = -0.3; increasing cos_sim to 0.9 makes it -0.4, lowering the loss. Take one gradient step on the embedding; with the printed sign the cosine similarity increases. Then inspect the authors' implementation (request code) to see whether the code uses `+ max(cos_sim - alpha, 0)`. If the code has the plus sign, Eqs. (3)-(4) are typos and the empirical claim is not yet invalidated but must be re-verified with corrected equations; if the code has the printed minus, the contrastive component as implemented would maximize cross-class similarity, so the reported FID gain cannot be attributed to the stated objective.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's second contribution is the contrastive loss in Eqs. (3)-(4). For a negative pair (different subclasses), the term is written as `- max((cos_sim - alpha), 0)`. If cos_sim > alpha, this term is negative and decreases as cos_sim increases, so gradient descent on L_cl pushes embeddings of different subclasses toward higher similarity. The stated goal—'minimizes the similarity between feature embeddings of ... different subclasses'—requires the opposite sign: `+ max((cos_sim - alpha), 0)` (or an equivalent hinge). This is not a cosmetic typo: the contrastive loss is one of the two named contributions, and Table 2 attributes the FID improvement partly to it. Because no code is provided, the reader cannot tell whether the implementation used the printed sign (in which case the method contradicts its own mechanism and the reported gains cannot be caused by the stated loss) or a corrected sign (in which case the paper's equations misrepresent the method). Either way the central claim is not verifiable from the manuscript. The label-supervision asymmetry noted by the reader is a real secondary concern, but this sign inconsistency is more load-bearing.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FG-RAT GAN, an extension of RAT GAN for fine-grained text-to-image synthesis. It adds an auxiliary classifier to the discriminator and a contrastive learning loss with a cross-batch memory (XBM) mechanism, and evaluates the resulting model on CUB-200-2011 and Oxford-102. The central claim, stated in the abstract and in Section 4.5, is that FG-RAT GAN achieves the lowest FID among the compared methods (LAFITE, VQ-Diffusion, RAT GAN) on both datasets while adding only a modest number of parameters.","tokens_in":12517,"tokens_out":4345,"duration_ms":52010,"significance":"If the reported results are valid, the paper would offer a lightweight and conceptually simple improvement over RAT GAN, with a parameter overhead of about 17M and a claimed FID reduction from 12.12 to 8.66 on CUB-200-2011 and from 12.90 to 9.14 on Oxford-102. The use of FID as the primary metric, with an explicit discussion of the limitations of Inception Score, is a reasonable methodological choice. However, the manuscript provides no code, no statistical uncertainty quantification, and the main comparison is not controlled for the extra class-label supervision used by the proposed method; these issues prevent the central claim from being verified as written. The auxiliary-classifier idea is well motivated, and the overall approach is easy to follow, but the load-bearing evaluation and the sign of the contrastive-loss term need correction before the contribution can be assessed.","major_comments":[{"comment":"The negative-pair term in the contrastive loss has the wrong sign relative to the stated objective. For a negative pair with cos_sim > alpha, the term is written as -max((cos_sim - alpha), 0), which is negative and decreases as cos_sim increases. Minimizing this term therefore encourages embeddings from different subclasses to become more similar, contradicting the text's claim that the loss 'minimizes the similarity between feature embeddings of ... different subclasses.' The correct hinge form should be +max((cos_sim - alpha), 0). Since the contrastive loss is one of the two named contributions and Table 2 attributes part of the FID improvement to it, this is not a cosmetic typo; as printed, the equations do not implement the described mechanism. The absence of released code makes it impossible to determine whether the implementation used the printed sign or a corrected one.","section":"Section 3.2, Eqs. (3)-(4)"},{"comment":"The comparison against LAFITE and VQ-Diffusion is not controlled for the additional supervision used by FG-RAT GAN. The proposed method trains an auxiliary classifier with ground-truth class labels, while the two baselines are label-free text-conditioned models. The sentence 'Even though we use labels during the training, label information is not an unfair advantage but a distinct characteristic of our model' is an assertion, not a controlled experiment. The ablation in Table 2 shows that RAT GAN + classifier, which also uses labels, already reduces FID from 12.12 to 9.90 on CUB, so the label signal alone explains a substantial part of the gain. A comparison with a label-conditioned baseline (e.g., ACGAN or TAC-GAN) or a version of the proposed model trained without the classifier is necessary to attribute the reported improvement to the proposed mechanisms.","section":"Section 4.5, Table 1"},{"comment":"All reported IS and FID values appear to come from a single training run, with no error bars, confidence intervals, or significance tests. The claimed gains are sometimes small (e.g., IS 5.08 vs 4.99 in Table 2 on CUB, and IS 3.66 vs 3.62 on Oxford), and FID is known to vary across seeds and evaluation batches. Without multiple runs and standard deviations, the reader cannot assess whether the reported FID differences, such as 8.66 vs 9.10 in Table 2, are statistically meaningful. The paper should report mean and standard deviation over at least three seeds for each configuration.","section":"Section 4.5, Tables 1 and 2"}],"minor_comments":[{"comment":"There are numerous typos and formatting errors, including 'Frenchet Inception Distance' (Section 4.2), 'horizontaly' (Section 4.1), 'orginal' (Section 4.4), 'auxliary' and 'discrminator' (Figure 1 caption), 'contrastive' misspelled as 'contrtastive' (Table 2), and a stray '0' at the end of Section 4.4. The manuscript should be carefully proofread.","section":"Throughout"},{"comment":"Reference [17] is the cross-batch memory (XBM) source, but reference [18] is a fine-grained image recognition paper by the current authors. Citing [18] together with [17] for the XBM mechanism is misleading unless the connection is explicitly explained; the citation should be corrected or removed.","section":"References, [17]-[18]"},{"comment":"The conclusion states that 'further evaluations on broader text-to-image synthesis benchmarks and more varied datasets are necessary to confirm the generalizability of our approach.' This is a self-acknowledged limitation, but the paper currently presents the method as achieving state-of-the-art performance; the scope of the claim should be aligned with this admitted limitation in the abstract and Section 4.5.","section":"Section 5, Conclusion"},{"comment":"The qualitative comparisons against DALL-E 2 and Stable Diffusion in the appendix are based only on a few cherry-picked examples and are not accompanied by any quantitative evaluation or human-study protocol. The text 'we demonstrate that our FG-RAT GAN can reach better visualized results compared with DALLE-2 and Stable Diffusion' overstates what such figures can support.","section":"Section 4.4 and Appendix 6.1"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eqs. (3)-(4) is the most serious technical issue; if the implementation matches the printed equations, the contrastive component would not do what the paper claims, and if the implementation uses the corrected sign, the equations misrepresent the method. Either way, the paper needs a substantial revision with code release or detailed training details, a controlled comparison that accounts for label supervision, and multi-seed evaluation. Given these issues, I cannot recommend acceptance in the current form, but the core idea is salvageable and the paper is within the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a small, honest-in-intent empirical extension of RAT GAN: add an auxiliary classifier to the discriminator and add a contrastive loss with a cross-batch memory. The best thing here is the ablation study (Table 2), which actually tries to separate the contributions of the classifier and the contrastive term, and the authors do acknowledge most of their building blocks (ACGAN, TAC-GAN, Ye et al., XBM). The reported FID improvements over RAT GAN are plausible for this kind of added supervision and the comparison to LAFITE and VQ-Diffusion is useful context even if not a controlled apples-to-apples test.\n\nThe soft spots are real, though. First, the label-supervision asymmetry: the method gets class labels, and the baselines in Table 1 mostly do not. The reader's take on this is fair; the authors' justification in Section 4.5 is not convincing. Second, there are no error bars or multiple seeds anywhere, so the headline FID numbers are single-run anecdotes. Third, the stress-test note is correct and it is worse than a typo: Equations 3 and 4 have the negative-pair term with a minus sign in front of the hinge. As printed, minimizing that term encourages higher cosine similarity for different classes, which is the opposite of the stated objective. The ablation shows a FID improvement when this contrastive loss is included, but if the equations are what was implemented, the mechanism contradicts itself; if the implementation used the corrected sign, the paper misrepresents the method. Either way the central claim about the contrastive contribution is not verifiable from the manuscript.\n\nThere is one citation-pattern oddity: reference [18] is cited as part of the XBM discussion, but that citation is the authors' own SwinTransFuse paper, which has nothing to do with cross-batch memory. That is a minor blemish, not evidence of a deeper problem.\n\nWho is this for? Someone working on GAN-based fine-grained text-to-image generation who wants a compact survey of the pieces and a plausible recipe. It is not a breakthrough; it is a workshop-level contribution with a fixable but currently invalidating flaw in the loss formulation. The label asymmetry and missing error bars would need to be addressed too.\n\nRecommendation: send it to a serious referee, because the combination is relevant and the ablation is a useful data point, but the referee should be told to check the sign of the contrastive loss and to request code before any acceptance. If the sign error is confirmed in the implementation, the paper needs major revision or withdrawal of the contrastive contribution.","headline":"A modest GAN-era fine-grained text-to-image extension whose FID gains are plausible but unverified, with a likely sign error in the contrastive loss equations.","tokens_in":12930,"tokens_out":655,"would_cite":false,"duration_ms":9564,"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":"This paper proposes FG-RAT GAN, which adds an auxiliary classifier and contrastive learning with cross-batch memory to RAT GAN, and reports the lowest FID scores on CUB-200-2011 and Oxford-102.","keywords":["fine-grained text-to-image synthesis","auxiliary classifier","contrastive learning","cross-batch memory","generative adversarial network","CUB-200-2011","Oxford-102","Fréchet Inception Distance"],"falsifier":"Train a version of RAT GAN that receives the same subclass labels as an additional input to the generator or discriminator but omits the auxiliary classifier and contrastive loss; if its FID matches the reported 8.66 on CUB-200-2011, the paper's attribution of the gain to the new losses is falsified. Repeating each configuration across multiple random seeds to obtain FID error bars would also test whether the reported gaps exceed run-to-run noise.","tokens_in":11950,"feed_emoji":"🖼️","tokens_out":4236,"duration_ms":35798,"temperature":0.7,"pith_summary":"This paper argues that the RAT GAN text-to-image model can be made to respect fine-grained class distinctions by adding two inexpensive components: an auxiliary classifier that forces the discriminator to predict the image's subclass, and a contrastive loss, backed by a cross-batch memory, that pulls same-class images together and pushes different-class images apart. On the CUB-200-2011 bird dataset and Oxford-102 flower dataset the augmented model, called FG-RAT GAN, reports the lowest Fréchet Inception Distance among LAFITE, VQ-Diffusion, and RAT GAN, meaning its images are statistically closer to real photos under the Inception feature metric. The paper treats this as evidence that fine-grained label information can be leveraged efficiently to improve synthesis quality.","feed_headline":"GAN with auxiliary classifier and contrastive loss sets new FID lows","feed_subtitle":"FG-RAT GAN reports FID 8.66 on CUB-200-2011 and 9.14 on Oxford-102, beating RAT GAN, LAFITE, and VQ-Diffusion.","key_machinery":"The mechanism is an auxiliary classifier plus contrastive embedding added to the RAT GAN discriminator. The auxiliary classifier computes a categorical cross-entropy loss that grades both real and fake images against subclass labels, feeding a gradient to the generator; the contrastive branch L2-normalizes 256-D embeddings and applies a margin-based cosine similarity loss against a cross-batch memory queue of real-image embeddings, so that same-class pairs are pulled together and different-class pairs are pushed apart past margin α. The two losses are summed with the adversarial loss after an initial phase that stabilizes the embeddings through classification alone.","core_discovery":"FG-RAT GAN extends RAT GAN's discriminator with a fully connected classification head trained with categorical cross-entropy on both real and fake images, and an embedding branch trained with a contrastive loss that compares current-batch features against a queue of past real-image features. The two losses join the original adversarial loss in both generator and discriminator updates. On CUB-200-2011 the model reports FID 8.66 versus 12.12 for RAT GAN, 10.32 for VQ-Diffusion, and 10.48 for LAFITE; on Oxford-102 it reports FID 9.14 versus 12.90 for RAT GAN. The paper's claim is that these gains come from the auxiliary classifier and contrastive learning making the generator produce images that are more internally consistent within a subclass and better separated across subclasses.","pith_inferences":["A natural stress test would run the same ablation with class labels removed from the auxiliary head entirely to see how much of the FID drop is attributable to label availability rather than to the contrastive geometry.","The cross-batch memory could be repurposed as a hard-negative mining mechanism for other conditional generators, not just GANs, by swapping the memory queue into diffusion or autoregressive training loops.","If label noise is introduced during training, the margin-based contrastive loss may degrade more gracefully than the hard classifier, a hypothesis the current single-run FID table does not test."],"forward_implications":["On both CUB-200-2011 and Oxford-102, FG-RAT GAN reports lower FID than RAT GAN, LAFITE, and VQ-Diffusion, so the added components transfer across two datasets with different class counts.","The full model beats both single-component ablations on FID, supporting the paper's claim that the classifier and contrastive learning are complementary rather than redundant.","Because only two fully connected layers (roughly 17M parameters) are added at training time, the method offers a parameter-efficient path to fine-grained text-to-image synthesis without scaling to billion-parameter autoregressive or diffusion models.","The generator's within-class similarity increases while cross-class similarity decreases, which is the geometric condition the contrastive loss directly enforces."],"supporting_citations":[{"why":"Supplies the RAT GAN backbone that FG-RAT GAN modifies.","marker":"[7]"},{"why":"Introduces the cross-batch memory queue used to collect hard negatives for contrastive loss.","marker":"[17]"},{"why":"Established auxiliary classifier GANs, the basis of the classification head.","marker":"[22]"},{"why":"Provides the contrastive learning formulation for text-to-image synthesis that this work extends.","marker":"[23]"},{"why":"VQ-Diffusion is a baseline whose reported FID scores are compared against.","marker":"[14]"},{"why":"LAFITE is a baseline whose FID score on CUB is the comparison target.","marker":"[5]"},{"why":"Defines the dataset splits used for evaluation.","marker":"[2]"}],"fun_headline_variants":["FG-RAT GAN: auxiliary classifier and contrastive loss improve FID","Contrastive learning and auxiliary classifier refine fine-grained GAN","New fine-grained GAN with classifier and contrastive loss cuts FID","Auxiliary classifier and contrastive loss boost fine-grained text-to-image","Auxiliary classifier and contrastive learning sharpen fine-grained GAN FID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that feeding subclass labels during training is not itself the cause of the FID improvement, so the gains can be attributed to the auxiliary classifier and contrastive learning rather than to the extra label signal.","fun_headline_variants_meta":{"raw":{"variants":["FG-RAT GAN: auxiliary classifier and contrastive loss improve FID","Contrastive learning and auxiliary classifier refine fine-grained GAN","New fine-grained GAN with classifier and contrastive loss cuts FID","Auxiliary classifier and contrastive loss boost fine-grained text-to-image","Auxiliary classifier and contrastive learning sharpen fine-grained GAN FID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3454,"prompt_tokens":935,"completion_tokens":2519,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":2423}},"tokens_in":551,"tokens_out":2519,"duration_ms":19688,"temperature":1.0,"reasoning_tokens":2423,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:01:30.062359+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a version of RAT GAN that receives the same subclass labels as an additional input to the generator or discriminator but omits the auxiliary classifier and contrastive loss; if its FID matches the reported 8.66 on CUB-200-2011, the paper's attribution of the gain to the new losses is falsified. Repeating each configuration across multiple random seeds to obtain FID error bars would also test whether the reported gaps exceed run-to-run noise.","supporting_citations":[{"cited_title":"In: IEEE Transactions on Multimedia, vol","cited_arxiv_id":null,"evidence_quote":"Supplies the RAT GAN backbone that FG-RAT GAN modifies."},{"cited_title":"In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp","cited_arxiv_id":null,"evidence_quote":"Introduces the cross-batch memory queue used to collect hard negatives for contrastive loss."},{"cited_title":"In: Proceedings of the 34th International Conference on Machine Learning (ICML’17), vol","cited_arxiv_id":null,"evidence_quote":"Established auxiliary classifier GANs, the basis of the classification head."},{"cited_title":"In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp","cited_arxiv_id":null,"evidence_quote":"VQ-Diffusion is a baseline whose reported FID scores are compared against."},{"cited_title":"In: 2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pp","cited_arxiv_id":null,"evidence_quote":"LAFITE is a baseline whose FID score on CUB is the comparison target."},{"cited_title":"In: Balcan, M.F., Weinberger, K.Q","cited_arxiv_id":null,"evidence_quote":"Defines the dataset splits used for evaluation."}],"review_version":1}