{"id":"c34f5e75-c206-45e7-9994-8be5f595d55b","arxiv_id":"2502.01507","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-embedding GAN with one generator-side and one discriminator-side text encoder, trained end-to-end, improves FID and R-precision over shared and pre-trained embeddings on three text-to-image benchmarks.","lead":"This paper trains two separate text embeddings inside a text-to-image GAN, one for realistic rendering and one for text-image alignment, and reports better FID and alignment on CUB, Oxford-102, and MS-COCO than shared or frozen embeddings. The design reduces parameter count, but its advantage over strong baselines on MS-COCO only appears after adding a gradient-penalty trick from prior work.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"R-precision is measured with the same D-side encoder that conditions generation, so dual-embedding alignment gains may be a metric artifact; an external-encoder evaluation is needed.","rationale":"The reader's conditional verdict identifies attribution as the weakest assumption and notes the R-precision circularity as a secondary weakness. I agree that the internal ablation supports the FID/IS benefit of decoupling generator-side and discriminator-side embeddings, but I would elevate the R-precision circularity to the primary load-bearing concern because the paper's headline claim about text-image alignment is directly scored by the very encoder used to condition generation. Table 6 makes the mechanism visible: the R-precision jump from row 3 to row 5 is driven by whether the generator sees SD, not by the separation of embeddings per se. The same issue weakens the cross-method alignment comparisons. This does not change the conditional verdict; it sharpens the condition: the alignment claim needs an external, frozen evaluation encoder to be accepted. I am not alleging any procedural misconduct; the protocol is conventional in the literature, but it is circular when the central claim concerns the benefit of a separately trained alignment encoder.","tokens_in":20298,"tokens_out":7862,"duration_ms":72173,"concrete_test":"Recompute R-precision for all variants in Tables 3 and 6 using a frozen external text encoder (e.g., the pre-trained DAMSM encoder from AttnGAN or a CLIP text encoder) and a fixed image feature extractor not trained inside DTE-GAN, keeping the generated images fixed. If the row-5 vs row-3 R-precision gap (86.64 vs 63.79) collapses below about 2 points, the dual embedding's alignment benefit is an artifact of conditioning on the evaluation encoder. If the gap persists, the alignment claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central alignment claim rests on R-precision, but the paper computes R-precision using its own discriminator-side Bi-LSTM encoder and discriminator image features (Appendix C.4). This is not an independent probe. In Table 6, R-precision jumps from 63.79 (row 3: generator does not receive SD) to 86.64 (row 5: generator conditions on SD) while the encoder used for retrieval is the same D-side encoder in both rows. The generator is trained to make images that maximize similarity to SD under exactly the model that later retrieves text for R-precision, so the metric rewards conditioning on the evaluation encoder rather than generic text-image alignment. The same circularity applies to Table 3 (DTE vs CLIP) and Table 4 (AttnGAN+DTE): the alignment scores are produced by an encoder trained inside the DTE pipeline. Consequently, the claim that the second embedding improves text-image alignment is not established; only the photo-realism/FID benefit of decoupling is supported by the internal ablation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DTE-GAN, an end-to-end text-to-image GAN that learns two separate Bi-LSTM text embeddings: a generator-side embedding optimized through the adversarial/generation losses and a discriminator-side embedding optimized through a multi-modal contrastive loss on real image-text pairs. The generator conditions on both embeddings, while the discriminator uses the discriminator-side embedding for contrastive alignment and for adversarial criticism. The paper reports IS, FID, and R-precision on CUB, Oxford-102, and MS-COCO, presents an internal ablation (Table 6) showing that the dual-embedding configuration outperforms shared-embedding variants, demonstrates integration into AttnGAN (Table 4), and shows reuse of the learned embeddings for text-guided image manipulation (Table 5).","tokens_in":20459,"tokens_out":6065,"duration_ms":50148,"significance":"If the central claim holds, decoupling a text representation optimized for photo-realism from one optimized for text-image alignment is a useful, parameter-efficient design principle for lightweight GAN-based text-to-image synthesis. The paper has real strengths: Table 6 is a controlled internal ablation that isolates the embedding organization, Table 4 tests portability to an existing architecture, Table 5 shows downstream reuse, and DTE-GAN uses far fewer parameters (11M) than DF-GAN and SSA-GAN. The main reservation is that R-precision is computed with the same discriminator-side encoder that the generator is trained to align with, so the alignment benefit attributed to the dual-embedding design is not independently confirmed. An external-encoder evaluation would substantially strengthen the core claim.","major_comments":[{"comment":"The manuscript computes R-precision using text features from the D-side Bi-LSTM sentence encoder and image features from the discriminator network (Appendix C.4). These are exactly the encoder and image feature extractor optimized by the contrastive losses in Eq. (6) and Eq. (11), and the generator is trained to maximize similarity of generated images to SD under this same encoder. Consequently, the R-precision numbers in Tables 1, 3, 4, and 6 partly measure whether the generator conditions on the evaluation encoder rather than generic text-image alignment. The CUB jump from 63.79 to 86.64 between Table 6 row 3 and row 5 illustrates this dependence. Please re-evaluate R-precision (or at least the main comparisons) with an external frozen text/image encoder, such as CLIP or a DAMSM encoder not trained inside the DTE-GAN pipeline, and report both sets of numbers. This is necessary to support the abstract's and conclusion's claims about text-to-image alignment.","section":"Appendix C.4; Eq. (6), Eq. (11)"},{"comment":"The abstract and Section 4.2 state that DTE-GAN reduces CUB FID 'from 14.06 to 13.67', but no method in Table 1 has a FID baseline of 14.06: TIME is 14.30, DF-GAN is 14.81, and DAE-GAN is 15.19. Please identify the correct baseline or correct the number, because as written the improvement claim is not verifiable from the reported tables.","section":"Section 1; Section 4.2; Table 1"},{"comment":"The paper attributes part of its cross-method superiority to the dual-embedding design, but the comparisons in Tables 1 and 2 involve models with different generators, discriminators, loss terms, and parameter counts; the manuscript itself notes that DTE-GAN halves the width of DF-GAN/SSA-GAN blocks. Table 6 varies only the embedding organization within DTE-GAN's own architecture, so it cannot separate the effect of the embedding scheme from the effect of the overall architecture in the Table 1 and Table 2 comparisons. Please either add architecture-matched baseline models with shared embeddings for the external comparisons or qualify the attribution accordingly.","section":"Section 4.2; Table 1; Table 6"},{"comment":"The sentence 'On MS-COCO, we achieve similar performance of DF-GAN and SSA-GAN with fewer parameters' is not supported by Table 1 unless it refers to DTE-GAN+MAGP: DTE-GAN itself reports COCO FID 25.17, while DF-GAN reports 19.32 and SSA-GAN reports 19.37; DTE-GAN+MAGP reports 19.69. Please correct the sentence or explicitly state that the comparable result is achieved by DTE-GAN+MAGP.","section":"Section 4.2; Table 1"}],"minor_comments":[{"comment":"The author line reads 'Y eruru Asrar Ahmed'; this appears to be a formatting artifact and should be corrected to 'Yeruru Asrar Ahmed'.","section":"Title page/author line"},{"comment":"There are numerous typos and garbled words, including 'covaraince', 'regualarisation', 'apporach', 'geandator', 'Futher', and 'perpal' in Figure 4; the manuscript needs a careful proofreading pass.","section":"Throughout"},{"comment":"The CLIP and CLIP+Gemb baselines are not fully specified: it should be stated explicitly whether they use the same DTE-GAN generator/discriminator architecture or a different GAN architecture, since otherwise the comparison is difficult to interpret.","section":"Table 3"},{"comment":"The method name 'MANIGAN' should be 'ManiGAN' for consistency with the referenced work.","section":"Table 5"},{"comment":"The caption 'Single embeddings with outLG training achieves superior performance' is grammatically unclear; it should likely read 'with LG training' or 'without LG training' as appropriate.","section":"Section 4.3.6/Table 8 caption"},{"comment":"Table 2 states that R-precision scores are not available in the literature for Oxford-102, yet Tables 3 and 6 report Oxford R-precision values for DTE and CLIP variants; this is not contradictory, but the manuscript should clarify that the literature comparison lacks such numbers while the paper independently reports them.","section":"Table 2; Table 3; Table 6"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is interesting and the controlled ablation is a genuine contribution, but the alignment metric issue is central to the main claim and should be resolved before publication. I would encourage the editor to request the external-encoder R-precision evaluation as a condition of revision, rather than treating it as optional."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea is simple and worth taking seriously: train two separate Bi-LSTM text encoders in a T2I GAN, one feeding the generator and optimized for image realism, the other feeding the discriminator and optimized for text-image alignment via contrastive loss, with no gradient flow between them. This is genuinely different from TIME, which uses a single end-to-end embedding, and from SD-GAN's Siamese structure. The ablation in Table 6 is the heart of the paper, and it holds up as an internal comparison: the full dual-embedding configuration with SD passed to the generator beats the shared-embedding variants on both FID and IS, not just on R-precision. The AttnGAN+DTE experiment also shows the idea transfers outside their own architecture. And they get these numbers with roughly half the parameters of DF-GAN or SSA-GAN.\n\nThe soft spots are real, though. Most importantly, R-precision is computed with the same discriminator-side Bi-LSTM and discriminator image encoder that the generator is explicitly trained to fool. In row 5 of Table 6, the generator conditions on SD and receives gradients from a contrastive loss against SD; the retrieval test then uses SD. That is circular. The stress-test note is correct: the 86.64 R-precision likely rewards conditioning on the evaluation encoder rather than generic text-image alignment. The authors need to report R-precision with an external encoder, such as CLIP, before claiming the second embedding improves alignment. The FID story is separate and still stands, because FID/IS use a fixed Inception network.\n\nOther issues: the main text says CUB FID drops from 14.06 to 13.67, but 14.06 appears in no table; the nearest listed baseline is TIME's 14.30. That looks like a typo but needs correcting. No code or trained models are released, which matters because the dual-caption and transformer-encoder results are only reported in the appendix. The cross-method tables compare against models with different generators and widths, so the dual-embedding design is not isolated by those numbers; the internal ablation is the only controlled evidence, which is reasonable but should be acknowledged.\n\nOverall, this is a useful design principle for lightweight T2I GANs, and the photo-realism benefit of decoupling is supported. The alignment benefit is not yet established. The paper deserves peer review, but the reviewers should ask for an external-encoder evaluation and artifact release before the numbers are taken as definitive. I'd send it to review.","headline":"Decoupling generator- and discriminator-side text embeddings is a plausible idea with a solid internal ablation, but the circular R-precision metric means the alignment claim doesn't hold up; the photo-realism gains do.","tokens_in":21046,"tokens_out":4202,"would_cite":false,"duration_ms":36698,"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":"Splitting text embeddings into a realism branch and an alignment branch improves text-to-image GANs.","keywords":["text-to-image synthesis","dual text embeddings","end-to-end training","generative adversarial networks","contrastive learning","photo-realism","text-image alignment"],"falsifier":"Run a strictly controlled comparison where the architecture is held fixed and only the embedding scheme changes—same generator and discriminator widths, losses, training schedule, and random seeds, with dual embeddings versus a single shared embedding—and test whether the dual version still improves FID and R-precision; alternatively, insert DTE-GAN's dual embeddings into DF-GAN's exact generator and discriminator and check whether FID improves over DF-GAN's reported 14.81 on CUB. If the gap disappears in either test, the dual-embedding claim is not responsible for the published advantage.","tokens_in":20045,"feed_emoji":"🎨","tokens_out":8304,"duration_ms":63214,"temperature":0.7,"pith_summary":"This paper proposes that text embeddings for text-to-image synthesis should be learned inside the synthesis network, rather than borrowed from a pre-trained, generically trained text encoder. Its concrete claim is that two separate text embeddings work better than one shared embedding: a generator-side embedding trained to make generated images photo-realistic, and a discriminator-side embedding trained to keep images aligned with their captions. The paper implements this as DTE-GAN, a single-stage GAN with two Bi-LSTM text encoders, and reports lower Fréchet Inception Distance on CUB and Oxford-102 than prior GAN approaches that use pre-trained embeddings, and comparable MS-COCO results with far fewer parameters. The authors also show the learned embeddings can be reused for text-to-image manipulation. If the attribution holds, the useful principle is to decouple 'make it look real' from 'match the text' as separate conditioning representations.","feed_headline":"Two text embeddings beat one for text-to-image GANs","feed_subtitle":"Splitting text encoding into realism and alignment branches improves FID on CUB, Oxford-102, and MS-COCO.","key_machinery":"The central object is the dual text embedding setup: two independent word-embedding layers ($W_G$ and $W_D$) with two Bi-LSTM sentence encoders, one producing the generator-side sentence embedding $S_G$ and the other the discriminator-side sentence embedding $S_D$. The mechanism that makes the split work is gradient isolation: $S_G$ learns only from generation losses (adversarial loss plus multi-modal contrastive loss on fake images), while $S_D$ learns only from the multi-modal contrastive loss on real image-text pairs, with a stop-gradient preventing noisy generation gradients from flowing into $S_D$ when it is passed to the generator.","core_discovery":"The central claim is that decoupling text conditioning into two independently trained embeddings improves text-to-image generation. DTE-GAN uses separate word embeddings and Bi-LSTM sentence encoders for the generator and the discriminator: the generator-side embedding is updated only by the adversarial loss and by a contrastive loss computed on generated images, while the discriminator-side embedding is updated only by a contrastive loss computed on real image-text pairs. The generator additionally receives a detached copy of the discriminator-side sentence embedding, which the authors call a 'sneak peek' at a stable, alignment-focused representation. The paper reports that a shared embedding trained with both losses produces worse images, and that feeding the noisy generator-side embedding to the discriminator degrades performance, which it reads as evidence that separate, purpose-specific embeddings are the effective design choice.","pith_inferences":["If the decoupling principle generalizes, frozen pre-trained text encoders like CLIP may be leaving generative performance on the table: they are optimized for alignment and retrieval, not for driving a decoder toward realism. A testable extension would be to add a trainable realism branch to a diffusion-based text-to-image model while keeping a frozen alignment branch.","The successful 'sneak peek'—feeding the generator a detached, contrastively trained sentence embedding—suggests a general design rule for conditional generative models: the cleanest representation of the condition is not necessarily the one that should be fine-tuned by the generator's own noisy gradients. This could be probed in other conditional GANs by adding or removing the stop-gradient on the","The same dual-embedding split could transfer to other multimodal generation tasks, such as text-to-video or layout-to-image generation, where a single conditioning embedding currently has to serve both realism and semantic fidelity."],"forward_implications":["On CUB, DTE-GAN lowers FID to 13.67 and raises R-precision to 86.64, beating DF-GAN (14.81 FID) and AttnGAN (23.98 FID) on the same dataset.","On Oxford-102, DTE-GAN lowers FID to 30.07 from DualAttn-GAN's 40.31 and raises Inception Score to 4.21.","On MS-COCO, DTE-GAN reaches FID 25.17 with 11M parameters, and 19.69 with DTE-GAN+MAGP, matching DF-GAN and SSA-GAN while using roughly half their parameter count.","Table 6 shows the dual-embedding configuration beats all shared-embedding variants of the same architecture on IS, FID, and R-precision.","The learned dual embeddings transfer: used as fixed word embeddings in a lightweight text-guided manipulation GAN, they improve FID from 8.02 to 7.77 on CUB."],"supporting_citations":[{"why":"Supplies the DAMSM pre-trained text embeddings and contrastive loss that DTE-GAN replaces, and the R-precision metric used for text-image alignment.","marker":"Xu et al., 2018"},{"why":"Provides CLIP, the representative pre-trained discriminative text encoder used in the DTE vs CLIP and CLIP+Gemb comparisons.","marker":"Radford et al., 2021"},{"why":"Defines the DF-GAN baseline and the Matching-Aware Gradient Penalty used in DTE-GAN+MAGP; DTE-GAN halves the width of its UpBlocks and DownBlocks.","marker":"Tao et al., 2022"},{"why":"Defines the SSA-GAN baseline on CUB and COCO that DTE-GAN+MAGP is compared against.","marker":"Liao et al., 2022"},{"why":"Supplies the conditioning augmentation technique used in DTE-GAN and the dataset setup for CUB and Oxford-102.","marker":"Zhang et al., 2017b"},{"why":"Defines the HDGAN baseline on Oxford-102 used for visual and quantitative comparison.","marker":"Zhang et al., 2018"},{"why":"Defines the Fréchet Inception Distance metric used to compare image quality across all experiments.","marker":"Heusel et al., 2017"}],"fun_headline_variants":["Dual text embeddings improve text-to-image generation","Two text embeddings beat a shared one in T2I GANs","Separate text embeddings for realism and alignment win","End-to-end dual text embeddings sharpen text-to-image GANs","Splitting text encoding improves photo-realism and alignment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion that the dual-embedding design itself causes the reported gains rests on ablations performed inside DTE-GAN's own architecture; comparisons against earlier methods change the generator, discriminator, and parameter count simultaneously, so the dual embeddings may not be the actual source of the improvement over those baselines.","fun_headline_variants_meta":{"raw":{"variants":["Dual text embeddings improve text-to-image generation","Two text embeddings beat a shared one in T2I GANs","Separate text embeddings for realism and alignment win","End-to-end dual text embeddings sharpen text-to-image GANs","Splitting text encoding improves photo-realism and alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1645,"prompt_tokens":931,"completion_tokens":714,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":632}},"tokens_in":547,"tokens_out":714,"duration_ms":6963,"temperature":1.0,"reasoning_tokens":632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:05:22.507560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a strictly controlled comparison where the architecture is held fixed and only the embedding scheme changes—same generator and discriminator widths, losses, training schedule, and random seeds, with dual embeddings versus a single shared embedding—and test whether the dual version still improves FID and R-precision; alternatively, insert DTE-GAN's dual embeddings into DF-GAN's exact generator and discriminator and check whether FID improves over DF-GAN's reported 14.81 on CUB. If the gap disappears in either test, the dual-embedding claim is not responsible for the published advantage.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the DF-GAN baseline and the Matching-Aware Gradient Penalty used in DTE-GAN+MAGP; DTE-GAN halves the width of its UpBlocks and DownBlocks."}],"review_version":1}