{"id":"ed57c682-530a-4e6d-8d85-83b014461935","arxiv_id":"2412.03957","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding label-based supervised contrastive losses to text-to-image GANs reduces FID by up to 30.1% on COCO and improves IS on CUB across four baselines.","lead":"This paper proposes adding a supervised contrastive loss, guided by class labels, to both the pretraining and GAN training phases of text-to-image GANs. The modified models report lower FID and higher IS than the original AttnGAN, DM-GAN, SSA-GAN, and GALIP on CUB and COCO.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported FID/IS gains likely depend on comparing against official pre-trained baselines rather than identically re-trained baselines; the paper does not state whether the 30.1% COCO AttnGAN FID improvement is against a re-trained baseline or the official model, and no code/checkpoints are…","rationale":"The reader's weakest_assumption focuses on label quality and the need for complete labels, a valid limitation but not the most load-bearing threat to the quantitative claim. The paper itself acknowledges the limitation, and COCO/CUB labels are detailed and known, so the label-fidelity assumption is likely satisfied. The more pressing issue is the comparability of the numbers: the paper says baselines come from publicly released pre-trained models ('*'), which means they are not re-trained in the same environment, while the +SCL results are new runs with different code, hardware, and potentially different data splits and evaluation details. Under these conditions, the reported FID/IS improvements may partly reflect re-training, hyperparameter tuning, or implementation differences rather than the added contrastive losses. The paper also omits error bars for FID, never reports multiple seeds, and does not release code or checkpoints. The central claim is plausible and the architecture is a straightforward extension, but the claimed magnitudes (especially 30.1%) are not adequately verified. A conditional acceptance with a request for re-run comparison, multiple seeds, evaluation protocol details, and code/checkpoints release is the appropriate verdict. The paper deserves credit for providing ablations and qualitative examples, but those do not settle the comparative question.","tokens_in":11685,"tokens_out":1604,"duration_ms":14501,"concrete_test":"Run the same training protocol in Table 3 of the official AttnGAN and DM-GAN repositories with identical data splits, batch sizes, seeds, and evaluation code (same FID feature set), re-training the baseline and the +SCL variant. If the baseline FID or the +SCL FID changes by more than ~5% relative to Table 1, or if the delta shrinks below the gap needed to support the claimed 30.1% improvement, then the headline claim is weakened. The paper would need to report the re-trained baseline numbers, multiple seeds, and the exact evaluation script.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that adding supervised contrastive losses improves FID/IS across four T2I GAN baselines, with the headline 30.1% FID reduction on COCO with AttnGAN. Table 1 marks baseline numbers with '*' as obtained from publicly released pre-trained models by the authors. This means the reported improvement mixes numbers produced by different training runs/setups, so the gain could partly be a re-training effect, not the proposed framework. The paper also gives no error bars for FID, no multiple seeds, and no code or checkpoint release. The ablation (Table 2) shows that pre-training alone improves AttnGAN COCO FID from 33.10 to 26.90 (18.7%), and the full method reaches 23.14 (30.1%), suggesting the comparison baseline may not match the official AttnGAN's own reported FID under identical evaluation conditions. The paper's claim that no existing approach leverages labels like ours is overstated given its own UniCL and cross-entropy comparisons, and the discussion in Sec. 5 acknowledges the method is only demonstrated on datasets with detailed label annotation. The load-bearing concern is not the method's validity but the empirical claim's comparative basis: it is unsupported without re-running baselines under identical data/seed/training settings, reporting variance, and releasing code.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework that adds label-guided supervised contrastive learning to text-to-image GANs in both the representation pretraining phase and the GAN training phase. The method samples same-label image-text pairs, computes supervised contrastive losses for image-image, text-text, and image-text associations (single-label for CUB, multi-label for COCO), and adds these to the original objectives with weights lambda1 and lambda2. The framework is applied to AttnGAN, DM-GAN, SSA-GAN, and GALIP, and the authors report improved IS and FID on CUB and COCO, with the headline result being a 30.1% FID improvement for AttnGAN on COCO. An ablation on AttnGAN decomposes the contribution of the pretraining component, the image-image loss, and the image-text loss.","tokens_in":11987,"tokens_out":5307,"duration_ms":46624,"significance":"If the empirical claims hold, the framework is a simple and general recipe for injecting label information into T2I GANs, and the choice of four baselines spanning multi-stage and one-stage architectures is appropriate. The loss definitions in Eqs. (1)-(9) are coherent, and the two-branch parameter-sharing design is clearly described. The paper also makes falsifiable quantitative predictions and includes an ablation study, which are strengths. However, the empirical verification is currently weakened by the use of public checkpoints as baselines rather than identically re-trained baselines, by the absence of variance or multiple-seed reporting for FID, and by the lack of released code or checkpoints. These issues bear directly on the central claim of large FID improvements, so the contribution is not yet fully established.","major_comments":[{"comment":"The note under Table 1 states that baseline results marked * are 'obtained from publicly released pre-trained models by the authors', while the '+SCL' results come from models newly trained by the authors. This mixes evaluation conditions: differences in data preprocessing, training epochs, random seeds, and evaluation code can produce FID changes of the magnitude claimed (e.g., 33.10 to 23.14 for AttnGAN on COCO). The central claim that the framework improves FID by 30.1% is therefore not supported unless each baseline is re-trained by the authors under the same pipeline (same data, same seeds, same evaluation script) and those re-trained baseline numbers are reported in Table 1.","section":"§4.1, Table 1"},{"comment":"No variance or number of seeds is reported for any FID score, and the IS values are reported with standard deviation but it is not stated over how many runs or seeds. GAN training is stochastic, and FID is known to vary across runs. Without at least three seeds and mean±std for both IS and FID, the reported improvements cannot be distinguished from run-to-run noise. This is particularly important for the smaller claimed gains (e.g., GALIP CUB FID 10.08 to 9.90).","section":"§4.1, Table 1"},{"comment":"The ablation shows that the pretraining component alone (ID 2) reduces AttnGAN COCO FID from 33.10 to 26.90, an 18.7% improvement, while the full method reaches 23.14. Because the ID 1 baseline is a public checkpoint rather than a model re-trained with the authors' data pipeline and training schedule, the improvement attributed to the GAN-phase losses (ID 3 vs ID 2, and ID 5 vs ID 2) is measured relative to a baseline that was not subject to the same pipeline. The ablation should include an AttnGAN baseline re-trained by the authors as ID 1, with all configurations trained for the same number of epochs, to isolate the effect of the proposed losses.","section":"§4.3, Table 2"},{"comment":"The values of lambda1 and lambda2 are set separately for each dataset and each baseline, and no sensitivity analysis is reported. Since these hyperparameters control the strength of the proposed losses, the reported gains could be affected by the choice of their values. Please provide at least a small sensitivity study (e.g., varying lambda2 for one baseline per dataset) or justify the chosen values by a fixed rule.","section":"§4 (Implementation Details)"}],"minor_comments":[{"comment":"The notation fYp, eYi, and eY is not explained; the symbols appear to mix a tilde or accent with Y. Please define Y_i, Y'_i, and the concatenated label set explicitly.","section":"§3.1, Eq. (1)"},{"comment":"In the definition of Pm(i), the condition should be that the labels have non-empty intersection, but the text writes 'fYp ∩ eYi' without explaining the intersection for the multi-label case. Please clarify the notation.","section":"§3.1, multi-label loss"},{"comment":"The statement that FID 'measures the KL divergence between the real images and generated images' is incorrect; FID computes the Fréchet distance between feature distributions. Please correct this explanation.","section":"§4.3"},{"comment":"The claim that no existing approach leverages labels like the proposed framework is contradicted by the UniCL and cross-entropy comparisons in the same section; please phrase the novelty claim more carefully.","section":"§4.4"},{"comment":"Reference [23] is cited for FID but points to 'Coulomb GANs' (Unterthiner et al.); the standard FID reference is Heusel et al., 'GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium'. Please fix the citation.","section":"References"},{"comment":"There are several typos, including 'Lorgin' in Eq. (5), 'vice verca' in Section 2.1, and 'noval' in Section 1. Please correct them.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a computer vision or generative modeling venue, and the method is plausible. The main weakness is empirical rigor: the comparison against public checkpoints, missing variance, and lack of code release prevent verification of the central claim. I would like to see the authors re-train baselines under identical conditions, report multiple seeds, and provide code or checkpoints before acceptance. The related-work novelty claim is somewhat overstated but can be fixed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean, incremental idea—label-supervised contrastive learning plugged into both the pretraining and GAN phases of T2I GANs—and the reported gains are plausible. But the empirical basis is weaker than it looks: the baseline numbers are from public checkpoints, not re-trained under the same conditions, so the headline FID drops partly reflect retraining effects, not just the SCL loss.\n\nWhat is new: applying SupCon to both phases with a two-branch, parameter-sharing design is a reasonable extension of Khosla et al. to T2I GANs. The paper compares against UniCL and cross-entropy, which is the right kind of control, and the ablation separates the contributions of the image-image and image-text losses. That is more than many papers in this area do. The writing is clear, the method is simple enough to apply to four different baselines, and the qualitative examples support the idea that same-label clustering helps.\n\nThe main soft spots: (1) the baseline issue above is load-bearing—the asterisk in Table 1 means the numbers are not apples-to-apples, and the stress-test note is right to flag it; (2) no code or checkpoints are released, so verification is impossible without reimplementation; (3) the claim that no existing approach uses labels like ours is overstated given UniCL, though they do compare with it; (4) the method needs detailed labels, which the authors acknowledge in Sec 5. These are real but not fatal—the ablation alone shows the pretraining step improves COCO FID by 18.7%, so the direction is believable. The precise magnitudes, especially the 30.1% AttnGAN gain, should be treated with caution until baselines are re-run and variance is reported.\n\nI don't see a circularity problem: the losses are standard, the metrics are external, and the comparison to UniCL and cross-entropy is fair. The paper is honest about its scope.\n\nWho should read this: people still working on GAN-based T2I generation, not the diffusion crowd. The practical impact is modest, but the idea is transferable. I would send it to review, with the clear expectation that the authors re-run baselines, report multiple seeds, and release code. The core idea is sound; the empirical claim needs to catch up.","headline":"A clean, incremental idea for T2I GANs whose reported gains are plausible but whose empirical basis is undermined by comparing against official checkpoints rather than re-trained baselines; deserves review but needs major revision.","tokens_in":12513,"tokens_out":3437,"would_cite":false,"duration_ms":32224,"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":"Label-guided supervised contrastive learning in both pretraining and GAN phases cuts COCO FID by up to 30.1% across four text-to-image GANs.","keywords":["text-to-image generation","GAN","supervised contrastive learning","semantic label guidance","image-text representation learning","multi-label contrastive learning","CUB dataset","COCO dataset"],"falsifier":"Train the framework on COCO with randomly permuted label sets, so that positives are assigned without regard to true semantics; if FID still improves by the reported margin over the unmodified baseline, then label-driven clustering is not the operative cause.","tokens_in":11502,"feed_emoji":"🎨","tokens_out":8857,"duration_ms":70573,"temperature":0.7,"pith_summary":"Typical text-to-image GANs align an image with its caption but ignore the semantic structure among images that share a category label. This paper proposes adding supervised contrastive losses, driven by image labels, to both the representation-pretraining phase and the GAN training phase, so that same-label image and text embeddings are pulled together while different-label pairs are pushed apart. On the COCO dataset the framework lowers FID by 30.1%, 27.3%, 16.2%, and 17.1% for AttnGAN, DM-GAN, SSA-GAN, and GALIP respectively, with smaller but consistent gains on CUB measured by IS and FID. The result is a training-time collaboration that improves generation quality, especially on complex multi-object scenes, without changing the base model architecture.","feed_headline":"Supervised contrastive learning cuts text-to-image FID by 30 percent","feed_subtitle":"By clustering same-label images and text, the method improves fidelity for four GAN baselines on CUB and COCO.","key_machinery":"The load-bearing object is the label-guided supervised contrastive loss\n$$\\mathcal{L}_{\\mathrm{sup}}(u_i,w)=-\\frac{1}{|P_s(i)|}\\sum_{p\\in P_s(i)}\\log\\frac{\\exp(\\mathrm{sim}(u_i,w_p)/\\tau)}{\\sum_{j\\ne i}\\exp(\\mathrm{sim}(u_i,w_j)/\\tau)},$$\nwhere $\\mathrm{sim}$ is cosine similarity and $P_s(i)$ collects indices whose labels match $u_i$'s label (or whose label sets intersect, in the multi-label COCO case). Around this loss the framework builds a data-sampling strategy that guarantees every sample has a same-label partner, and twin parameter-sharing encoder branches that produce quadruples of image and text representations for both the pretraining phase and the GAN phase. In the GAN phase the same loss is applied to fake images and their captions, so the generator is rewarded for emitting images that cluster with other images of the same category.","core_discovery":"The paper's central claim is that inner-modal semantic correspondence—the fact that images with the same label share visual patterns—is an underused signal in text-to-image GANs, and that supervised contrastive learning can convert that signal into better generation. The framework samples, for each image-text pair, a partner pair sharing at least one label, forms two parameter-sharing contrast branches, and adds three supervised contrastive losses ($\\mathcal{L}_{\\mathrm{sup}}^{\\mathrm{img}}$, $\\mathcal{L}_{\\mathrm{sup}}^{\\mathrm{txt}}$, $\\mathcal{L}_{\\mathrm{sup}}^{\\mathrm{i2t}}$) during pretraining, plus image-image and image-text versions during GAN training. With single-label positives on CUB and multi-label positives on COCO, the authors report that AttnGAN's IS rises from 4.36 to 4.61 and its COCO FID falls from 33.10 to 23.14, and every tested baseline improves on both datasets; the largest relative gains appear on the more complex COCO dataset.","pith_inferences":["A direct extension the authors leave to future work is to replace human labels with nouns extracted from captions; the multi-label contrastive loss could absorb such pseudo-labels, and the method's success would then depend on the precision of the extracted positives.","The same loss could likely be inserted into diffusion-based text-to-image pipelines, since it operates on encoder outputs and generated images rather than on any GAN-specific objective; the authors mention diffusion only as a possible future direction.","The framework's clustering pressure may trade fidelity for diversity: as the authors note, adding the image-text loss can lower CUB FID, and a dedicated intra-class diversity metric would reveal whether same-label positives reduce variety within a category."],"forward_implications":["The framework is architecture-agnostic: applying it to multi-stage GANs (AttnGAN, DM-GAN) and one-stage GANs (SSA-GAN, GALIP) improves FID on COCO and CUB in every case.","The ablation attributes the gain to all three components: improved pretrained encoders alone take AttnGAN's COCO FID from 33.10 to 26.90, and adding the GAN-phase image-image loss yields the largest additional drop.","The relative benefit is larger on multi-object COCO (up to 30.1% FID reduction) than on single-object CUB, consistent with the claim that label structure is more informative when scenes are complex.","Because the pretrained encoders are frozen in the GAN phase, the additional machinery is limited to the two-branch sampling and contrastive losses on generated quadruples.","Using the image-image and image-text supervised contrastive losses together gives the best overall result, although adding the image-text loss alone can slightly worsen CUB FID relative to image-image alone."],"supporting_citations":[{"why":"Defines the single-label supervised contrastive loss that the pretraining objective adapts.","marker":"[6]"},{"why":"Provides the multi-label contrastive loss used for COCO's overlapping label sets.","marker":"[12]"},{"why":"Is the AttnGAN baseline and the source of the DAMSM pretraining objective reused for three of the baselines.","marker":"[25]"},{"why":"Is the DM-GAN baseline whose dynamic-memory generator is extended with the framework.","marker":"[32]"},{"why":"Is the SSA-GAN baseline that supplies the one-stage architecture and the note that IS is unreliable on COCO.","marker":"[9]"},{"why":"Is the GALIP baseline that supplies CLIP-based encoders and the branch structure on which the framework is applied.","marker":"[21]"},{"why":"Provides the CLIP encoders and symmetric cross-entropy loss used when the baseline is GALIP.","marker":"[15]"},{"why":"Is the UniCL baseline the paper compares against as an alternative label-guidance method.","marker":"[26]"},{"why":"Is the COCO dataset whose multi-label annotations drive the multi-label positive sampling.","marker":"[11]"},{"why":"Is the CUB dataset whose single-label annotations drive the single-label positive sampling.","marker":"[24]"}],"fun_headline_variants":["Supervised contrast learning cuts COCO FID by up to 30%","Inner-modal contrastive training enhances text-to-image GANs","Label-guided contrast branches improve GAN image synthesis","Same-label clustering sharpens images in four GAN baselines","Contrastive learning with label info reduces T2I errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that every image has accurate, complete category labels and that label overlap is a trustworthy signal of semantic similarity, so noisy, missing, or overly coarse labels would create false positive pairs and could degrade rather than improve generation.","fun_headline_variants_meta":{"raw":{"variants":["Supervised contrast learning cuts COCO FID by up to 30%","Inner-modal contrastive training enhances text-to-image GANs","Label-guided contrast branches improve GAN image synthesis","Same-label clustering sharpens images in four GAN baselines","Contrastive learning with label info reduces T2I errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000595,"raw_usage":{"total_tokens":2851,"prompt_tokens":1079,"completion_tokens":1772,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":1686}},"tokens_in":695,"tokens_out":1772,"duration_ms":12265,"temperature":1.0,"reasoning_tokens":1686,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:53:53.942141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the framework on COCO with randomly permuted label sets, so that positives are assigned without regard to true semantics; if FID still improves by the reported margin over the unmodified baseline, then label-driven clustering is not the operative cause.","supporting_citations":[{"cited_title":"IEEE Transactions on Neural Networks and Learning Systems (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the multi-label contrastive loss used for COCO's overlapping label sets."},{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Is the AttnGAN baseline and the source of the DAMSM pretraining objective reused for three of the baselines."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Is the DM-GAN baseline whose dynamic-memory generator is extended with the framework."},{"cited_title":"In: Meila, M., Zhang, T","cited_arxiv_id":null,"evidence_quote":"Provides the CLIP encoders and symmetric cross-entropy loss used when the baseline is GALIP."},{"cited_title":"In: Computer Vision– ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13","cited_arxiv_id":null,"evidence_quote":"Is the COCO dataset whose multi-label annotations drive the multi-label positive sampling."},{"cited_title":"california institute of technology (2011)","cited_arxiv_id":null,"evidence_quote":"Is the CUB dataset whose single-label annotations drive the single-label positive sampling."}],"review_version":1}