{"id":"f8e256dd-6b73-4197-a441-6aad67755b35","arxiv_id":"2504.17990","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A two-stage framework, TSCIR, separates learning image-to-pseudo-word mapping from learning to compose those words with modification text, achieving state-of-the-art zero-shot composed image retrieval on Fashion-IQ, CIRR, and CIRCO.","lead":"This paper introduces a two-stage training method for zero-shot composed image retrieval, where an image is first converted into a pseudo-word token and then the model learns to combine that token with text. It reports state-of-the-art results on three public benchmarks using far less synthetic training data than prior methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage II's Laion-CIR training triples are not checked for overlap with evaluation galleries; since CIRCO uses COCO images that web-crawled LAION very likely contains, the 3.58-point CIRCO mAP@5 gain over Stage I could be image-identity leakage rather than compositional learning.","rationale":"I agree with the reader's conditional verdict. The paper's strongest claim—competitive with large-scale synthetic methods and SOTA with 32K synthetic triplets—rests on benchmark numbers whose validity depends on train/test image disjointness. The paper explicitly asserts this only for CC3M in Section IV-A. Because CIRCO is COCO-based and LAION is a web crawl with known COCO presence, the missing overlap audit is not a formality; it is the difference between 'composes' and 'recognizes training images.' The two-stage design itself is coherent and the ablations support the component-level claims (VSI, soft alignment, CAs, hard negatives), so I would not reject. But the current evidence does not yet establish that the Stage II gains on CIRCO—and possibly CIRR/FashionIQ—are compositional. Code is not released and no error bars are reported, which makes the overlap check even more important. A secondary description issue (no explicit exclusion of the ground-truth target in hard-negative generation, Eqs. 21–22) is worth fixing but is less load-bearing, since the reported positive effect of HNL implies the implementation likely excludes or handles it. Overall, the reader's CONDITIONAL recommendation is the right one; my stress-test does not move it.","tokens_in":15289,"tokens_out":5716,"duration_ms":60320,"concrete_test":"Compute exact/near-duplicate overlap between every image in the 32K Laion-CIR synthetic training set (reference and target sides) and the gallery images of CIRCO, CIRR, and Fashion-IQ, using a combination of URL/ID matching and perceptual hashing (e.g., dHash/pHash) with manual inspection of near-duplicate pairs. Report the overlap count per evaluation set. If the count is nonzero (especially for CIRCO), retrain the Stage II adapters on the deduplicated subset and re-run Tables I–III; if CIRCO mAP@5 falls from 18.37 to near the 14.8–15.1 range, the headline two-stage gain is dominated by leakage. If the overlap count is zero, the concern is resolved and the current SOTA claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that Stage II turns a strong mapping (Stage I) into a state-of-the-art composer using only 32K synthetic triplets (TSCIR 32k rows in Tables I–III). The load-bearing premise is that those synthetic triplets do not leak evaluation images. Section IV-A verifies non-overlap only for CC3M ('no overlap with the evaluation datasets'), not for the Laion-CIR-Template/Laion-CIR-LLM data. This matters disproportionately for CIRCO, which is constructed from COCO images; LAION-400M is a web-scale crawl and is known to contain COCO/Flickr images, so CIRCO reference/target images are plausibly present in the 32K training set. If present, the frozen image encoder plus the 1.18M adapters can learn to associate a composed text embedding with a specific training image's feature, and that image will rank high at test time when it appears in the CIRCO gallery. The magnitude of the reported gain is consistent with this worry: CIRCO mAP@5 rises from 14.79 (Stage I) to 18.37 (Stage II), and mAP@50 from 19.00 to 22.71, a much larger jump than the difference among published synthetic-data methods (e.g., FTI4CIR 15.05 mAP@5). A second, smaller description gap is that Eqs. (21)–(22) do not state that the ground-truth target is excluded from the top-k candidate set; if it is included, β=1 makes the positive also a negative. The reported improvements suggest the implementation handles this, but the text should say so. The primary concern is testable and does not require doubting the architecture; it requires a data-hygiene audit before the SOTA claim is accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TSCIR, a two-stage zero-shot composed image retrieval framework. Stage I trains an image-to-pseudo-word mapping on CC3M using a visual semantic injection (VSI) module and a soft text alignment loss; Stage II uses roughly 32K synthetic triplets (Laion-CIR from TransAgg) to train lightweight composing adapters and a hard-negative interpolation loss while freezing the backbone. The method reports state-of-the-art results on Fashion-IQ, CIRR, and CIRCO, with ablations showing that each component contributes and that Stage I alone is competitive with methods trained on large-scale synthetic data.","tokens_in":15672,"tokens_out":8120,"duration_ms":72017,"significance":"The central idea—decoupling mapping from composing and showing that Stage II needs only a small amount of synthetic data—is timely and, if the results are clean, would be a useful step for zero-shot CIR. The paper's strengths include systematic ablations (Tables IV–VIII) that support the incremental contribution of VSI, soft text alignment, adapters, and hard negatives, and the use of public benchmarks and public synthetic data, which aids reproducibility. However, the headline numbers depend on data hygiene and on a few technical statements that need correction; these issues must be resolved before the state-of-the-art claim can be accepted.","major_comments":[{"comment":"The paper states in Section IV-A that CC3M has 'no overlap with the evaluation datasets,' but it does not report a similar check for the Laion-CIR synthetic triplets used in Stage II. Because CIRCO is built from COCO images and LAION is a web-scale crawl likely to contain COCO images, the Stage II CIRCO gains (Table III: mAP@5 from 14.79 to 18.37, mAP@50 from 19.00 to 22.71) could partly reflect retrieval of training images by identity rather than compositional understanding. Please add exact/near-duplicate overlap statistics between the Laion-CIR training images and the CIRR/CIRCO/Fashion-IQ galleries, and either retrain on the non-overlapping subset or show that removing overlaps does not change the conclusions.","section":"Section IV-A / Table III"},{"comment":"The soft text alignment loss is not well defined as written. Equation (10) defines P as a negative batch-averaged scalar softmax value, and Eq. (11) similarly defines Q, so P and Q are negative scalars, not probability distributions; the KL divergence in Eq. (12) is therefore undefined. Please rewrite the loss as a per-sample KL between per-sample softmax distributions P_i and Q_i (presumably P_i = exp(τ vg_i^T cg_i)/Σ_j exp(τ vg_i^T cg_j) and Q_i = exp(τ sg_i^T cg_i)/Σ_j exp(τ sg_i^T cg_j), averaged over i), and verify that the reported implementation matches this corrected form.","section":"Section III-A, Eqs. (10)–(12)"},{"comment":"The text in Section IV-C says the method 'consistently achieves superior results across multiple benchmarks,' but Table I shows CompoDiff (18.8M synthetic triplets) reports FashionIQ Avg R@10 = 36.02, above TSCIR-32K's 31.88, even though TSCIR has higher R@50. Please report the missing per-category numbers for CompoDiff, or qualify the comparison to the specific metrics/categories where the claim holds. The abstract's 'superior performance compared to existing approaches' should be made precise.","section":"Section IV-C / Table I"},{"comment":"The hard-negative interpolation in Eqs. (21)–(22) does not state that the ground-truth target t_gt is excluded from the top-k candidate set T. If t_gt is among the top-k, then its similarity with itself yields β=1 by Eq. (22), so the generated 'negative' equals the positive target. Please state explicitly that t_gt is removed from T before computing the mixing weights, or revise the formula to guarantee that generated negatives are distinct from the positive.","section":"Section III-B, Eqs. (21)–(22)"}],"minor_comments":[{"comment":"The text says the adapter configuration is placed at positions {2,4,6,8,10}, but Table VIII only reports {2,4,6,8,10,12}; the text and the table should be aligned.","section":"Section IV-B / Table VIII"},{"comment":"There are typos: 'FishionIQ' in Section IV-A and Figure 4 should be 'FashionIQ', and 'Synthetic daa' in the Table III caption should be 'Synthetic data'.","section":"Section IV-A / Table III caption"},{"comment":"Equation (19) uses h_g without defining it; please clarify whether h_g is the final adapted composed text feature and how it relates to z_c^g in Eq. (14).","section":"Section III-B, Eq. (19)"},{"comment":"The phrase 'text encoder with the V A' appears to be a typo for 'text encoder with VSI and CA'.","section":"Section IV-B"},{"comment":"The caption says 'visual semantic inject module'; this should be 'visual semantic injection module'.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the absence of overlap analysis for Laion-CIR relative to the evaluation galleries; if the authors cannot establish non-overlap, the CIRCO numbers in particular lose evidential value for compositional understanding. The remaining technical issues—the soft-alignment loss formula, the hard-negative exclusion step, and the overbroad SOTA claim—are fixable in revision, but all should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The two-stage decomposition is the real contribution here: first train the image-to-pseudo-word mapping, then teach the text encoder to compose that pseudo-word with modification text using only 32K synthetic triplets. That is genuinely different from prior work, which either trains only the mapping or jointly trains mapping and composing on millions of synthetic samples. The ablations are consistent, and each component—VSI, soft text alignment, composing adapters, hard negatives—pulls its weight. I buy the architecture and the design reasoning.\n\nThe soft spot is the CIRCO result. The synthetic training data comes from LAION, and CIRCO is built from COCO images. LAION is a web-scale crawl and is known to contain COCO. The paper only claims non-overlap for CC3M, not for the synthetic triplets. If any of those LAION images are in the CIRCO gallery, Stage II can memorize them and the mAP gain could be image-identity retrieval, not compositional understanding. That would inflate the headline numbers. This is testable and the authors should report the overlap explicitly.\n\nThe hard-negative construction in Eqs. (21)-(22) has a similar clarity gap: it doesn't say the ground-truth target is excluded from the top-k candidate set. If it's not, the positive becomes its own negative during training. The reported gains suggest the implementation handles this, but the text should be unambiguous. Also no code and no error bars, which is common but still makes the numbers hard to evaluate precisely.\n\nThese issues don't sink the central idea. The two-stage framework is a useful contribution regardless of the exact CIRCO number, and the ablations give me confidence the method does what it claims. The FashionIQ comparison to CompoDiff is a little selective—CompoDiff has higher R@10, though TSCIR wins on average—but that's a footnote, not a fatal issue.\n\nI'd want the authors to run the data hygiene audit and fix the hard-negative description before publication. If the overlap checks out, this is solid work. Even if some overlap exists and the CIRCO numbers drop, the two-stage design is still worth publishing. Send it to peer review, but ask for those checks.","headline":"A plausible two-stage idea with strong ablations, but the CIRCO numbers need a data hygiene audit before you trust the SOTA claim.","tokens_in":16248,"tokens_out":4222,"would_cite":true,"duration_ms":40058,"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 claims that decoupling zero-shot composed image retrieval into a visual-mapping stage and a composing stage, trained on only 32K synthetic triplets, beats prior methods on Fashion-IQ, CIRR, and CIRCO.","keywords":["zero-shot composed image retrieval","pseudo-word token","visual semantic injection","soft text alignment","composing adapter","hard negative mining","synthetic triplets","CLIP"],"falsifier":"Run a near-duplicate or exact-duplicate image search between the Laion-CIR synthetic training set (and the sampled SynTriplets18M subset) and the CIRCO, CIRR, and Fashion-IQ evaluation images. If even a small percentage of training reference/target images match evaluation query/target images, the Stage-II margins—e.g., CIRCO mAP@5 rising from 15.05 to 18.37—can be explained by memorized identity; if no matching pairs are found, the compositional interpretation survives.","tokens_in":15077,"feed_emoji":"🖼️","tokens_out":5865,"duration_ms":52961,"temperature":0.7,"pith_summary":"Zero-shot composed image retrieval asks a model to find a target image from a reference image plus a short modification text, without human-annotated retrieval triplets. This paper tries to show that the usual single-stage recipe—train an image-to-pseudo-word mapping, then hope the text encoder can combine that word with the modification—is the bottleneck, and that splitting training into two stages fixes it. In Stage I the model learns a rich image-to-pseudo-word mapping using ordinary image-caption pairs; in Stage II it learns to compose that pseudo-word with modification text using only about 32K synthetic triplets. If right, the result matters because it would slash the need for large synthetic datasets while improving retrieval accuracy.","feed_headline":"Zero-shot image retrieval: two-stage training with 32K triplets wins","feed_subtitle":"A mapping stage plus a composing stage on small synthetic data beats methods trained on millions of triplets.","key_machinery":"The load-bearing mechanism is the separation of mapping from composing. Stage I uses a Visual Semantic Injection (VSI) module—cross-attention layers that inject projected image patch features into the pseudo-word token inside the CLIP text encoder—plus a soft text alignment loss that pushes the pseudo-word's text features toward caption features according to a similarity distribution. Stage II freezes the backbone and adds lightweight composing adapters (two MLPs with residual connections, about 1M parameters) so the text encoder can combine the pseudo-word with modification text, trained with contrastive losses and a hard-negative strategy that interpolates target features with top-k distractors. The design keeps visual information concentrated in the pseudo-word token rather than flooding all words, which the ablations show is critical.","core_discovery":"The paper's central claim is that its two-stage framework, TSCIR, reaches state-of-the-art zero-shot composed image retrieval on Fashion-IQ, CIRR, and CIRCO, including CIRCO mAP@5 of 18.37 against 15.05 for the strongest prior method and Fashion-IQ average recall@10 of 31.88 against 30.61 for the fully fine-tuned TransAgg baseline. Even without Stage II, its Stage I mapping already outperforms all non-synthetic baselines and rivals methods trained on large synthetic sets. Using the same 32K synthetic triplets as TransAgg, TSCIR beats it by roughly 11.6% relative average recall on Fashion-IQ, and it stays competitive with methods using far larger synthetic corpora. The authors attribute this to a pseudo-word token that carries richer visual semantics plus a text encoder explicitly tuned to compose that token with real words.","pith_inferences":["A testable extension: if the two-stage separation is the operative factor, replacing CC3M with any domain-matched image-caption set in Stage I should transfer TSCIR to new retrieval domains with no triplet annotations.","The hard-negative mixup strategy is not limited to CIR; it could be isolated and tested in other metric-learning settings where hard negatives are rare.","The authors' implicit claim that pseudo-word quality, not composition machinery, drives most of the gain should be probed by scaling the backbone: ablations with ViT-H improve over ViT-L, so tests at larger scale would show whether composition learning grows with capacity.","The CIRCO numbers deserve an overlap audit before use in leaderboards, since CIRCO is built from COCO while Laion-CIR is web-harvested from a large corpus that contains COCO-derived images."],"forward_implications":["With the same 32K synthetic triplets as TransAgg, TSCIR reports about 11.6% relative improvement in average recall on Fashion-IQ, so synthetic-data efficiency is a direct corollary rather than a side effect.","Stage I alone beats every non-synthetic baseline on all three benchmarks, implying the mapping stage is independently valuable even when no synthetic triplets are available.","Adding Stage II to Stage I improves results on Fashion-IQ, CIRR, and CIRCO, so the two objectives are complementary rather than redundant.","Because the same framework gains from both diffusion-generated SynTriplets18M and retrieval-based Laion-CIR, the approach does not depend on one particular synthetic-data quality level."],"supporting_citations":[{"why":"Provides the CC3M image-caption pairs used to train the Stage I visual-to-pseudo mapping.","marker":"[10]"},{"why":"Pic2Word defines the projection-based baseline that maps global image features to a pseudo-word token.","marker":"[11]"},{"why":"Context-I2W is the token-learner baseline whose context-dependent mapping TSCIR extends.","marker":"[12]"},{"why":"CLIP supplies the visual and text encoders used as the backbone.","marker":"[15]"},{"why":"CompoDiff contributes the SynTriplets18M synthetic dataset and the large-scale-synthetic baseline.","marker":"[18]"},{"why":"Fashion-IQ is one of the three evaluation benchmarks.","marker":"[21]"},{"why":"CIRR is the real-life open-domain evaluation benchmark.","marker":"[22]"},{"why":"SEARLE-XL is a non-synthetic baseline and defines CIRCO, the third benchmark.","marker":"[23]"},{"why":"TransAgg provides the 32K Laion-CIR synthetic triplets used in Stage II and the direct same-data baseline.","marker":"[41]"}],"fun_headline_variants":["Two-stage zero-shot CIR: mapping then composing with 32K triplets","Small synthetic data, big retrieval gains: two-stage zero-shot CIR","CIR from mapping to composing: 32K triplets rival millions","Zero-shot CIR: two-stage mapping-composing SOTA on three datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central load-bearing premise is that the 32K synthetic Laion-CIR triplets used in Stage II do not materially overlap with the evaluation images; the paper verifies training/evaluation disjointness only for CC3M in Stage I, so if Laion-CIR contains near-duplicates of CIRCO/COCO images, the reported Stage-II gains could be identity retrieval rather than compositional understanding.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage zero-shot CIR: mapping then composing with 32K triplets","Small synthetic data, big retrieval gains: two-stage zero-shot CIR","CIR from mapping to composing: 32K triplets rival millions","Zero-shot CIR: two-stage mapping-composing SOTA on three datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3818,"prompt_tokens":997,"completion_tokens":2821,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":2739}},"tokens_in":613,"tokens_out":2821,"duration_ms":21297,"temperature":1.0,"reasoning_tokens":2739,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:28:22.139015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a near-duplicate or exact-duplicate image search between the Laion-CIR synthetic training set (and the sampled SynTriplets18M subset) and the CIRCO, CIRR, and Fashion-IQ evaluation images. If even a small percentage of training reference/target images match evaluation query/target images, the Stage-II margins—e.g., CIRCO mAP@5 rising from 15.05 to 18.37—can be explained by memorized identity; if no matching pairs are found, the compositional interpretation survives.","supporting_citations":[{"cited_title":"Pic2word: Mapping pictures to words for zero-shot com- posed image retrieval,","cited_arxiv_id":null,"evidence_quote":"Pic2Word defines the projection-based baseline that maps global image features to a pseudo-word token."},{"cited_title":"Context- i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval,","cited_arxiv_id":null,"evidence_quote":"Context-I2W is the token-learner baseline whose context-dependent mapping TSCIR extends."},{"cited_title":"Compodiff: Versatile composed image retrieval with latent diffusion,","cited_arxiv_id":null,"evidence_quote":"CompoDiff contributes the SynTriplets18M synthetic dataset and the large-scale-synthetic baseline."},{"cited_title":"Fashion iq: A new dataset towards retrieving images by natural language feedback,","cited_arxiv_id":null,"evidence_quote":"Fashion-IQ is one of the three evaluation benchmarks."},{"cited_title":"Zero-shot composed image retrieval with textual inversion,","cited_arxiv_id":null,"evidence_quote":"SEARLE-XL is a non-synthetic baseline and defines CIRCO, the third benchmark."}],"review_version":1}