{"id":"0acb3396-0327-4e26-afe1-2373d2f52c02","arxiv_id":"2411.16096","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"ENCLIP uses an ensemble of epoch-staggered CLIP fine-tunes plus K-means clustering to rank fashion search results, reporting gains over CLIP and FashionCLIP on a 44k-image dataset.","lead":"The paper proposes ENCLIP, a method that fine-tunes several CLIP models for different numbers of epochs, combines their outputs with a weighted score, and uses K-means clustering on t-SNE features to rank fashion image search results. The authors report improved mAP@10 over pretrained CLIP and FashionCLIP on a small low-resolution fashion dataset, but the evaluation is missing key baselines and the algorithm is not fully specified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 cannot produce the reported text-to-image retrieval numbers because it never takes a text query as input, and the reported gains over FashionCLIP are not attributable to the ensemble as described.","rationale":"The reader's weakest_assumption is exactly the load-bearing gap I find: the described algorithm has no text input, yet it is used to support text-to-image retrieval claims. I independently confirm that Algorithm 1's Require line lists only image outputs and N, and no later step embeds or injects the query; Section 4.4 discusses ranked images and clusters but never defines a text-image similarity. This is not merely a presentation issue: without a defined query pathway, the reported mAP numbers cannot be reproduced from the paper, and the comparison in Table 3 cannot be attributed to ENCLIP as specified. Additional weaknesses that reinforce rejection include the absence of a single fine-tuned CLIP control, test-set-dependent selection of weights and ranking criteria, and no variance estimates. I agree with the reader's verdict of REJECT with moderate confidence; the gaps are severe enough that the central empirical claim is unverified, though there is no evidence of fabrication and the conclusion itself concedes limitations for fine-grained queries. A concrete re-derivation and baseline test would settle the matter, but as the paper stands it does not meet the bar for acceptance.","tokens_in":8091,"tokens_out":1435,"duration_ms":12386,"concrete_test":"Ask the authors to provide the exact inference-time procedure executed for Table 3: the text encoder used, the query preprocessing, the similarity metric, and precisely which lines of Algorithm 1 use the query embedding. Then re-run mAP@10 on the 100 test queries with that procedure. Separately, fine-tune a single CLIP model for 80 or 100 epochs on the same 35k training set and compare its mAP@10 against ENCLIP; if the single fine-tuned model matches the ensemble, the claimed ensemble/clustering advantage is not supported.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that ENCLIP outperforms pretrained CLIP and FashionCLIP on text-to-image fashion search (mAP@10). However, Algorithm 1 (Section 4.4) has no text input: its inputs are image outputs M={m1..mz} and N, and its outputs are ranked images. The paper never specifies how a query string such as 'Give me polo neck t-shirt for men' is embedded, how that embedding is compared to the image embeddings, or how the algorithmic steps (frequency over model outputs, K-means on t-SNE, head-cluster selection) are combined with that similarity. Figure 4 states that retrieval uses the ENCLIP algorithm after storing embeddings, but the route from query text to ranked images is undefined. Additionally, the only quantitative comparison (Table 3) pits ENCLIP against zero-shot FashionCLIP and pretrained CLIP, not against a single fine-tuned CLIP baseline, so the ensemble's claimed contribution is not isolated. The weighting rule (Eq. 3), the ranking criterion (frequency and weighted_score), and the cluster count k in [4,6] are selected after observing the test queries, and no error bars are given. As written, the paper does not provide a reproducible procedure that generates the reported mAP values.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ENCLIP, an ensemble of five CLIP models fine-tuned for 10, 30, 50, 80, and 100 epochs on the Fashion Product Images (Small) dataset, combined through a weighted frequency score and K-means cluster-head selection. The authors claim that ENCLIP improves text-to-image fashion search over pretrained CLIP and FashionCLIP, as measured by mAP@10 across ten fashion categories. The manuscript includes qualitative retrieval examples, t-SNE visualizations, and a comparison table; it does not release code or specify how text queries enter the proposed algorithm.","tokens_in":8462,"tokens_out":3432,"duration_ms":33348,"significance":"If the central claim were supported, ENCLIP would offer a practical, low-resource recipe for adapting CLIP to small and low-quality fashion collections, using about 35k low-resolution images instead of the roughly 700k high-resolution images used by FashionCLIP. The paper also clearly identifies a real domain problem and reports results on a public dataset. However, the evidence as presented does not establish the claim: Algorithm 1 does not take a text query as input, the quantitative comparison does not include a single fine-tuned CLIP baseline, the relevance labels used for mAP are undefined, and key hyperparameters appear to be selected on the evaluation queries. These gaps concern the central claim and currently prevent the paper from being a reproducible or convincing empirical study.","major_comments":[{"comment":"Algorithm 1 is presented as the ENCLIP retrieval method, but its inputs are only image outputs M = {m1,...,mz} and an integer N; no text query is provided. The paper reports text-to-image retrieval results for queries such as \"Give me polo neck t-shirt for men\" (Tables 1 and 2 and Figures 5 and 6), yet it never specifies how the query string is embedded, how its embedding is compared with image embeddings, or how the frequency/weighted-score/clustering steps interact with that comparison. As written, the algorithm cannot produce the reported retrieval numbers, so the central claim lacks a defined mechanism.","section":"Section 4.4, Algorithm 1"},{"comment":"The only quantitative comparison is against zero-shot pretrained CLIP and zero-shot FashionCLIP, not against a single CLIP model fine-tuned on the same training data. Because ENCLIP is fine-tuned and the baselines are zero-shot, the reported gains could be due entirely to fine-tuning rather than to the ensemble or clustering strategy. The paper needs matched baselines such as individually fine-tuned CLIP models at each epoch count, and ideally an ablated ENCLIP without clustering, to isolate the contribution of the proposed components.","section":"Table 3 and Section 5"},{"comment":"The mAP computation relies on rel(k), the relevance function, but the paper never defines relevance labels for the Fashion Product Images (Small) dataset or describes how ground truth is constructed for the 100 evaluation queries. Without a definition of what makes a retrieved image relevant to a query, the mAP numbers in Table 3 are not interpretable or reproducible. This is a load-bearing omission for an empirical retrieval paper.","section":"Equation (6) and Section 5"},{"comment":"The cluster count k in [4,6] and the ranking criterion \"frequency and weighted_score\" are chosen after observing performance on the evaluation queries (\"The K value between the range of 4 and 6 gave the best results\" and \"Ranking based on frequency and weighted_score yielded the best results\"). No validation set or error bars are reported. This post-hoc selection on the test queries makes the reported mAP values optimistic and prevents a fair comparison with baselines that do not benefit from such selection.","section":"Section 4.4 and Section 5"},{"comment":"The training loss is described as the average of a categorical cross-entropy image loss and a categorical cross-entropy text loss, rather than the contrastive image-text loss used by CLIP. This is inconsistent with the paper's claim of fine-tuning CLIP and raises the question of what objective the individual models actually optimize. The authors should clarify whether they replace CLIP's contrastive objective and, if so, how the learned representations retain the cross-modal alignment needed for text-to-image search.","section":"Section 4.3, Equation (1)"}],"minor_comments":[{"comment":"Equation (2) in Section 4.4 and Equation (3) in Algorithm 1 are identical and both numbered; one of the two should be removed to avoid confusion.","section":"Equations (2) and (3)"},{"comment":"There are several typographical issues, including \"In 2023 Min Wang etal.\" missing a space, \"FashionCLIP, being a CLIP-like model,require\" missing an 's', and inconsistent spacing around commas and parentheses.","section":"Throughout"},{"comment":"The captions say the t-SNE plots and K-means results are \"obtained by the query\", but Algorithm 1 does not use a text query; the figures should either describe the actual selection process or be updated after the algorithm is revised to include query input.","section":"Figures 5 and 6"},{"comment":"Clustering is applied to t-SNE-transformed features, but t-SNE is stochastic and primarily preserves local neighborhoods; the authors should justify why clustering in t-SNE space is preferable to clustering in the original 512-dimensional latent space.","section":"Section 4.4"},{"comment":"The table reports only point estimates of mAP@10 without standard deviations or per-query breakdowns; at minimum, the paper should report error bars over multiple runs or query subsets, especially because the ensemble contains stochastic components.","section":"Table 3"}],"recommendation":"reject","confidential_remarks":"The stress-test concern about Algorithm 1 genuinely lands: the paper's central text-to-image retrieval claim cannot be produced by the algorithm as stated. The comparison is also not apples-to-apples because FashionCLIP and pretrained CLIP are used zero-shot while ENCLIP is fine-tuned, and relevance labels are never defined. These are fixable in principle, but the required experiments and algorithmic specification are substantial enough that the manuscript is not ready for publication in its current form. I would not recommend transferring the paper to another venue without the missing evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe gist: this is an engineering report on ensembling CLIP fine-tunes at different epochs plus K-means clustering for fashion retrieval on a small, low-res dataset. The combination is new as a package, and the problem is real. But the paper as written does not support its central claim.\n\nWhat it does well: it identifies a real practical issue—CLIP fine-tuning on ~35k low-res fashion images—and the idea of blending models at different training stages has face validity. The qualitative example (polo neck t-shirt) shows the ensemble output looks more on-target than the zero-shot baselines. The paper also honestly notes its own limitation with fine-grained queries.\n\nThe soft spots are load-bearing. Algorithm 1, the ENCLIP algorithm, takes no text query as input. Its inputs are the image outputs of the five fine-tuned models and a number N. It ranks images by frequency and weighted score, independent of any query. Yet the paper reports text-to-image search results. Either the algorithm description is missing a retrieval step, or the reported numbers cannot come from the described method. The paper never says how a query string is embedded and combined with the ranking. That is an internal contradiction, not a minor omission.\n\nSecond, Table 3 compares ENCLIP against zero-shot pretrained CLIP and zero-shot FashionCLIP. There is no single-fine-tuned-CLIP baseline. So even if the numbers are reproducible, you cannot attribute the gains to the ensemble or clustering. Third, the weighting rule (0.1 x 2^n), the cluster count k in [4,6], and the ranking criterion were chosen after looking at test-query performance. That is post-hoc fitting. And there are no error bars, no definition of relevance labels, and no code.\n\nThe upshot: the paper is a promising engineering sketch, not a defensible experimental claim. A rewrite that specifies the full pipeline, adds a fine-tuned single-model baseline, and reports variance could make it a useful workshop paper. As it stands, the central claim is unsupported.\n\nRecommendation: desk reject for now; the authors should be encouraged to resubmit with the missing details.","headline":"Useful problem and a plausible ensemble idea, but Algorithm 1 has no text input and there is no fine-tuned baseline, so the reported text-to-image numbers do not stand.","tokens_in":8857,"tokens_out":2315,"would_cite":false,"duration_ms":21955,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ENCLIP claims that ensembling five CLIP models fine-tuned for different epoch counts improves fashion text-to-image search on small, low-quality datasets beyond pretrained CLIP and FashionCLIP.","keywords":["multimodal search","information retrieval","fashion e-commerce","ensembling","clustering","CLIP","contrastive learning","low-resolution images"],"falsifier":"Re-run ENCLIP exactly as described on the same dataset with the same 100 queries and compare the resulting mAP@10 values with Table 3; if the procedure cannot accept a text query or the numbers do not match, the central claim is falsified. A smaller check: the stated score gives zero weight to the first fine-tuned model, so removing it should not change the ranking if the formula is applied literally.","tokens_in":7880,"feed_emoji":"👗","tokens_out":9127,"duration_ms":76445,"temperature":0.7,"pith_summary":"This paper claims that a single fine-tuned CLIP model is not the best way to adapt CLIP to small, low-quality fashion image collections, and that ensembling several fine-tuned CLIP models can do better. It proposes ENCLIP, which trains five CLIP models for 10, 30, 50, 80, and 100 epochs, encodes the dataset images with each, and then selects retrieval results by weighted frequency scores plus K-means cluster-head ranking. On the Fashion Product Images (Small) dataset, ENCLIP reports higher mean average precision at 10 (mAP@10) than both pretrained CLIP and FashionCLIP for most of the tested query categories. If the result holds, it gives a practical recipe for squeezing better text-to-image search out of CLIP when training data are scarce and images are low resolution, as is common in fashion e-commerce.","feed_headline":"Five fine-tuned CLIP models beat FashionCLIP on fuzzy fashion search","feed_subtitle":"Weighted ensemble plus cluster-head ranking lifts mAP@10 on a 44k low-resolution fashion set.","key_machinery":"The central object is the ENCLIP selection algorithm (Algorithm 1), a weighted-voting and cluster-head ranking procedure. Each fine-tuned CLIP model outputs a latent image embedding; occurrences of an image across models define a frequency, and the weighted score is $\\sum_{n=0}^{z-1} 0.1(2n)\\times\\text{occurrence}$, which gives later training stages greater influence. K-means is applied to the t-SNE-reduced latent space to define clusters, and the most frequent image in each cluster acts as the head image whose cluster is then ranked by frequency and weighted score. The mechanism's job is to let early-epoch and late-epoch checkpoints complement each other: early checkpoints generalize on scarce data, late checkpoints capture finer distinctions, and the cluster step suppresses images that only one model accidentally retrieved.","core_discovery":"ENCLIP is an ensemble-and-cluster method for fine-tuning CLIP in a data-poor, low-image-quality setting. Five CLIP models are initialized identically and trained for different numbers of epochs, and each model's image outputs are treated as votes. An image's frequency across model outputs and a weighted score $0.1 \\times (2n)$ per occurrence at model stage $n$ determine priority, and K-means clustering in a t-SNE-reduced latent space selects head clusters from which the top $N$ images are ranked. The paper reports that this combined ranking gives mAP@10 values above pretrained CLIP in all ten category-subcategory cells and above FashionCLIP in eight of ten cells, with the largest gains on ethnic wear for men and women's footwear. The intended message is that early-epoch models generalize and late-epoch models are precise, so their ensemble plus clustering is a better fit for limited-data fashion search than any single checkpoint.","pith_inferences":["A reader of Algorithm 1 will notice that its inputs are image outputs and a number $N$, with no text query; since the paper reports text-to-image mAP@10, there must be an unreported step that embeds the query and combines it with the image-based selection, and that step should be made explicit before the results are reproduced.","The weighting formula assigns zero weight to the first model's occurrences, so one testable implication is that dropping the epoch-10 checkpoint should not change the ranking; if it does, the formula in the paper is not the one actually evaluated.","The same ensemble-and-cluster procedure could be applied to other contrastive image-text models and to other small, noisy image domains once the query-fusion step is specified.","The evaluation uses only 10 queries per category, so a larger query set would test whether the reported margins are stable."],"forward_implications":["The reported mAP@10 places ENCLIP above pretrained CLIP in all ten category-subcategory cells and above FashionCLIP in eight of ten cells for the 100 queries tested.","The approach achieves these numbers with roughly 35k low-resolution training images, about 20 times fewer than the high-resolution set used to train FashionCLIP, so the ensemble-and-cluster recipe is presented as a substitute for a large clean dataset.","The authors report the largest relative gains in categories where pretrained CLIP is weakest, such as ethnic wear for men, so ensembling is claimed to recover signal in underrepresented fashion segments.","Because the method only adds standard CLIP fine-tuning runs, t-SNE projection, and K-means, it can be applied to an existing small catalog without collecting new images or changing the model architecture."],"supporting_citations":[{"why":"Supplies the Fashion Product Images (Small) dataset of roughly 44k low-resolution image-text pairs used for fine-tuning and evaluation.","marker":"[1]"},{"why":"Supplies the mAP and precision@k formulas used to score retrieval results.","marker":"[4]"},{"why":"Defines the FashionCLIP baseline and the high-resolution training regime that ENCLIP compares against with far fewer low-resolution images.","marker":"[5]"},{"why":"Defines CLIP, the contrastive image-text model that ENCLIP fine-tunes in five epoch variants.","marker":"[15]"},{"why":"Provides the elbow and silhouette methods used to choose the number of K-means clusters in the selection step.","marker":"[20]"}],"fun_headline_variants":["Ensemble of five CLIP models beats FashionCLIP on fuzzy fashion search","CLIP ensemble and clustering outperform FashionCLIP on limited data","Five CLIP checkpoints ensembled with clustering beat FashionCLIP on fuzzy data","Cluster-boosted CLIP ensemble wins on fashion search with low-quality images","Grouped CLIP checkpoints beat single models on low-res fashion search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central assumption is that its described ENCLIP procedure is what produced the reported retrieval numbers; the listed algorithm takes no text query, so the step that connects a user's text to the image selection is unstated.","fun_headline_variants_meta":{"raw":{"variants":["Ensemble of five CLIP models beats FashionCLIP on fuzzy fashion search","CLIP ensemble and clustering outperform FashionCLIP on limited data","Five CLIP checkpoints ensembled with clustering beat FashionCLIP on fuzzy data","Cluster-boosted CLIP ensemble wins on fashion search with low-quality images","Grouped CLIP checkpoints beat single models on low-res fashion search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000691,"raw_usage":{"total_tokens":3140,"prompt_tokens":969,"completion_tokens":2171,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":2072}},"tokens_in":585,"tokens_out":2171,"duration_ms":14242,"temperature":1.0,"reasoning_tokens":2072,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:32:46.855554+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run ENCLIP exactly as described on the same dataset with the same 100 queries and compare the resulting mAP@10 values with Table 3; if the procedure cannot accept a text query or the numbers do not match, the central claim is falsified. A smaller check: the stated score gives zero weight to the first fine-tuned model, so removing it should not change the ranking if the formula is applied literally.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the mAP and precision@k formulas used to score retrieval results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Fashion Product Images (Small) dataset of roughly 44k low-resolution image-text pairs used for fine-tuning and evaluation."},{"cited_title":"J., Attanasio, G., Bianchi, F., Terragni, S., Magalhães, A","cited_arxiv_id":null,"evidence_quote":"Defines the FashionCLIP baseline and the high-resolution training regime that ENCLIP compares against with far fewer low-resolution images."},{"cited_title":"W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., & Sastry, G","cited_arxiv_id":null,"evidence_quote":"Defines CLIP, the contrastive image-text model that ENCLIP fine-tunes in five epoch variants."},{"cited_title":"M., Saputra, D., & Oswari, L","cited_arxiv_id":null,"evidence_quote":"Provides the elbow and silhouette methods used to choose the number of K-means clusters in the selection step."}],"review_version":1}