Instance-Level Composed Image Retrieval

We thank Reviewer FuSR for emphasising the key strengths of our work:
(i) a carefully engineered, extensible i-CIR benchmark,
(ii) a genuinely training-free BASIC pipeline that unlocks pre-trained VLMs for CIR, and
(iii) consistently promising results across five datasets.
Below we clarify the remaining concerns—most notably hyper-parameter (HP) selection—and address every question in turn.

Hyper-parameter Tuning is Never Done on the Test Sets

We apologize for any confusion and we would like to thank the reviewer for this comment. We have omitted some information in the original submission, which we provide herein. During the development of the BASIC method, we reserved a small slice of the i-CIR crawl as a development set; none of those images appear in the final test benchmark. i-CIR_{dev} consists of 15 object instances, 92 composed queries, and 45K images in total. All BASIC HPs—contrastive scaling α, number of PCA components k, and Harris regularization λ—were fixed on this split and then frozen. The purpose of Fig. S3 is to show that BASIC’s performance curve is stable over a reasonably wide neighbourhood of hyperparameters (HP). To further demonstrate this robustness, we now perform a cross-dataset check. We set each dataset in turn as the development set and evaluate on the others in a “leave-one-out” logic.
Even though the largest performance appears on the diagonal, there is no large difference compared to tuning on the development dataset (or any other dataset). A small exception is LTLL, presumably due to its limited unique textual queries (2) and its very small size.
We shall definitely add this leave-one-out table in the supplementary material of the paper. We shall also explicitly mention the development set that was used for tuning BASIC's HP.

``Learned projection'' vs Training-Free

We use the term “learned” as commonly used for PCA which is data-dependent.
Therefore, the projection in BASIC is not learned via supervision or optimization, but rather computed analytically from two fixed text corpora: (C_+) (object-centric terms) and (C_-) (style/context terms). As described in section 4 (Eq. 2), we construct a contrastive covariance matrix and perform eigendecomposition to extract the top positive eigenvectors, which define the projection space. This process involves no training, backpropagation, or labeled data, and is consistent with BASIC’s training-free design. To avoid misinterpretation, we will revise the sentence.

LAION Statistics Do Not Leak Test Information

LAION is very large; i-CIR overlaps with it by <0.001 %. Moreover, each instance of i-CIR has its own tailored database consisting of the explicitly curated hard negatives. Using the global CLIP LAION mean therefore does not leak any test set statistics, yet gives a domain-agnostic center that transfers well across all datasets. To substantiate this, we computed the center on ImageNet-1K and simply replaced the LAION mean with this new one. BASIC on i-CIR achieves 46.67% mAP with the ImageNet mean versus 47.39% mAP with the LAION mean, a negligible drop of <0.8 mAP. This small gap confirms that our results are not artefacts of test leakage, but are the result of the BASIC design.

Details of i-CIR Construction

The labelling of the neighbours retrieved from LAION is a pseudo-labelling. For example, if the seed sentence is a textual description of the instance like “Temple of Poseidon”, then the retrieved neighbours should depict exactly this instance OR in a worse case just “temples”. Thus, these images are pseudo-labeled/classified as visual hard negatives. This is what L126 “for cases (i & ii), (ii), and (iv & v), respectively” was trying to describe. We shall rephrase for more clarity. The annotators were trained workers performing a well-defined annotation task. They completed a one-week training on composed image retrieval task, as also on copyright and sensitive-content removal. We shall include these details in the supplementary material. Thanks for the suggestion.

Why BASIC Trails in the Fashion Category

MagicLens is pre-trained on 36.7 M triplets dominated by words such as “brand” and “style” (according to their Fig. 5 b), which explains its narrow advantage in fashion. Outside that niche, BASIC leads by up to 34.8 % mAP (household category). We will add exemplar failure cases per visual category in the supplementary material.

Other Questions & Promised Edits

We will consider a more specific variant of the title. We will move some info from the i-CIR comparison (current section 3.4) into section 2 and add some comments positioning BASIC among composed image retrieval methods. We will correct any typos.

We thank Reviewer pyz4 for the review and for explicitly recognising several core strengths of our submission:
(i) a clear motivation punctuated by Figs. 1 & 3 and the full annotation pipeline in section 3,
(ii) instance-level CIR is a real-world task,
(iii) BASIC is a training-free method with a concise and efficient fusion strategy, especially suitable for scenarios with low resources or high deployment flexibility requirements,
(iv) strong, broad empirical results across i-CIR, ImageNet-R, MiniDN, and LTLL,
(v) ablations in Table 2 verify the logic and value of each component of BASIC.

Adaptability of a Fixed, Training-Free BASIC

BASIC purposefully avoids task-specific learning so that it inherits CLIP’s open-domain coverage instead of narrowing it down. BASIC relies on manually designed components, each motivated by intuitive principles: semantic projection isolates object-level semantics, Harris fusion balances dual-modality relevance, and contextualization mitigates CLIP’s sensitivity to phrasing. These components are carefully chosen to remain training-free and interpretable, achieving SOTA performance across diverse and shifted domains and showing strong robustness without requiring adaptation. Nonetheless, as the reviewer rightly notes, BASIC also inherits CLIP’s limitations—especially in the textual domain, where phrasing, abstraction level, or relational structure can affect performance. For visual inputs, CLIP’s vast pre-training set offers strong coverage. While fine-tuning could improve in-domain performance, it risks reducing generality.

A more targeted and compatible extension would be to explore hybrid text embedding strategies. For example, training a lightweight text encoder for relational or instructional queries, related to the following reviewer’s comment, or fine-tuning a contextualization-like embedding to reduce sensitivity and inference overhead. These directions remain promising extensions that target specific constraints (e.g., CLIP sensitivity to query wording or time overhead of contextualization).

Longer, Relational, or Instructional Text Queries

Thank you for the comment, which led us to the following insightful analysis. We assume that a text query can be almost equivalently formulated in different ways, e.g. concisely (“engraving”), in a longer sentence (“as a vintage stylish engraving”), in a relational way (“the same landmark as the one depicted in the image but as an engraving”), and in an instructional way (“engrave the landmark depicted in the image”). In the context of i-CIR, we refer to those cases as original, longer, relational and instructional. The queries of i-CIR are brief and comprehensive, while existing datasets like CIRR and CIRCO are often relational or/and instructional, while having larger lengths.

To probe BASIC’s robustness against such text query styles we automatically rewrite (via ChatGPT) all text queries of i-CIR, evaluate mmAP and present the results in the Table below (Δ is the relative change with respect to the original row). This is an automated process that allows us to perform such an insightful experiment despite the possibility of noise. Additionally, we perform the reverse mapping for 10 relational queries of CIRR that are converted into shorter and absolute descriptions. Results are discussed below.

Longer. We expand all text queries by adding 1–4 extra words while preserving their meaning. On average the phrases grow from 4.4 → 7.9 words—an increase of 3.5 words or +79 %, effectively almost doubling their length. The median rises from 4 to 8 words, the shortest query goes from 1 to 4 words, and the longest from 13 to 15. Example: “crowded” becomes “in a dense and crowded scene”. The ChatGPT prompt used to convert the text queries is:

Take each text query and, without altering its core meaning or introducing new concepts, enrich it with tasteful descriptive words or clarifying phrases so that its length is roughly doubled. Preserve the original order and casing and return each pair as ‘original query’ : ‘augmented longer query’ with no additional commentary. Example: ‘crowded’ → ‘in a dense and crowded scene’

We observe that BASIC has a small drop in performance similar to that of FreeDom, while MagicLens has a larger drop, therefore revealing a new advantage of training-free (relying on CLIP) methods. Related to the above question, MagicLens presumably has a larger drop because of shorter form sentences in its training.

Relational. To give every text query an explicit link to its paired reference image, we rewrite all text queries by prepending a short relational clause while keeping the original order and casing intact. Example: “with fog” becomes “the same object instance depicted in the image but with fog”. This yields a suite of explicitly relational captions that stress compositional understanding without altering the original intent. The ChatGPT prompt used to convert the text queries is:

Rewrite every text query so that it explicitly refers to the same object instance of the image query, while preserving original words in the same order. Simply prepend a concise relational clause, avoid introducing new concepts or proper names, and return each pair as ‘original query’ : ‘relational query’, with no additional commentary. Example: ‘with fog’ becomes ‘the same object instance depicted in the image but with fog’

We observe that the drop of BASIC is noticeably smaller than that of FreeDom, but also larger than that of MagicLens. This pronounces the benefit of training for composed image retrieval, and a drawback of training-free methods, which is due to the fact that there is no or little relational information during the CLIP pre-training. Nevertheless, BASIC still performs reasonably well.

Interestingly, evaluating using the 10 rewritten queries of CIRR (shown in the Table below) to make them non-relational (concise) demonstrates a large performance increase for BASIC from 18% mAP (10% R@1) to 44.4% mAP (40% R@1). This signifies that the relational description might not be the best way to describe those queries either.

Instructional. We recast every text query as an imperative instruction that tells the system what to do with the reference-image object. Concretely, we prepend an action verb (e.g. “turn”, “place”, “render”) and then a clause like “the object depicted in the image”, then append the original words in the original order. Example: “in a live action scene” becomes “adapt the object depicted in the image into a live action scene”. The ChatGPT prompt used was:

Rewrite each text query as a concise imperative instruction that starts with an appropriate action verb, explicitly mentions something like ‘the object depicted in the image’, preserves the original words in the same order, introduces no new concepts, and output each pair exactly as ‘original query’ : ‘instructional query’ with no commentary.

The performance pattern across the methods in this case mirrors the relational case.

Newer Baselines

The experiments already include comparisons against methods from 2024-2025. Specifically, FreeDom is a WACV 2025 oral, MagicLens is an ICML 2024 oral, CIReVL (supplementary Table S2) is an ICLR 2024 poster publication. Following the suggestion of Reviewer rzCM, we have now included CoVR-2 (TPAMI 2024) and MCL (ICML 2024). BASIC consistently outperforms both methods across 5 CIR datasets.

Robustness of Projection & Fusion Corpora

As described in supplementary § S2, we use ChatGPT to construct a generic object-centric corpus (C_{+}) that defines the projection subspace. The corresponding stylistic/contextual corpus (C_{-}) is detailed in our response to Reviewer przL. These corpora are fixed and shared across all datasets and domains, enabling training-free and scalable retrieval without task-specific tuning. To assess robustness, Tables S5 and S6 of the supplementary report results using domain-specific corpora (e.g. mobility-, landmark-, or fashion-focused). These dedicated variants yield only a +0.4 mAP improvement on average, indicating that the generic corpus is already effective across diverse datasets. At the same time, the projection matrix can be recomputed from any text pool in under 1 second on CPU, making the method trivially adaptable at deployment time if domain-specific knowledge is available. A natural direction for future work is to automatically construct such corpora by parsing the dataset or its metadata using a large vision-language model (LVLM), which could eliminate the need for manual corpus design entirely.

I thank the authors for their detailed responses to my concerns. The main confusion I raised: how the model handles longer, more compositional or relational queries with fine-grained semantic understanding, is now perfectly explained. After re-evaluation, I will increase my rating accordingly.

We thank Reviewer rzCM for a very positive assessment and for highlighting that:
(i) the i-CIR construction pipeline is transparent and well-motivated,
(ii) the work is backed by a comprehensive experimental suite and rich supplementary material that provide strong empirical evidence for every claim,
(iii) the manuscript is exceptionally clear—with intuitive figures (particularly the section 4 pipeline) and plentiful illustrative examples—making the narrative easy to follow, and
(iv) the combination of an ambiguity-free benchmark and a simple yet effective training-free BASIC method represents a significant and original step forward for late-interaction approaches in composed image retrieval.
Below we address the remaining suggestions and concerns:

Comparison with Additional SOTA Training-Based Methods

We appreciate the reviewer’s suggestion to include recent training-based methods. Using the authors’ official code and checkpoints, we evaluated CoVR-2 and MCL on the five CIR benchmarks of Table 1 of the main paper. As shown in the table below, our training-free BASIC still achieves the best mAP on every dataset—often by a sizeable margin (e.g. +20.6% on ImageNet-R and +10.5% on i-CIR). These results strengthen the evidence that BASIC offers competitive accuracy without the computational cost of training or fine-tuning.

Why Gains Are Smaller on CIRR, FashionIQ, and CIRCO

1: These three benchmarks suffer from modality domination: many queries are answerable from the text alone. To make the point quantitative, we swept a mixing weight $λ ∈ [0,1]$ between text-only $λ = 0$ and image-only $λ = 1$ similarity for three naïve fusion methods (WeiCom, Sum, and Product). The average (over three fusion methods) peak–vs.–unimodal gap is a good proxy for how much a dataset rewards composition. On i-CIR the average gap is enormous: +16.1 mAP (+391% relative), whereas it shrinks to +3.0 mAP (+11%) on CIRR, +5.0 mAP (+26%) on FashionIQ, and +6.8 mAP (+167%) on CIRCO. Moreover, the highest unimodal performance is always when $λ = 0$, i.e. text-only. Thus legacy datasets reward composition only marginally, whereas i-CIR demands strong cross-modal synergy; BASIC is designed for the latter scenario. Detailed results are presented on the Table below.
2: Sensitivity to relational or instructional phrasing (link to Reviewer pyz4). As shown in our answer to Reviewer pyz4, rewriting just ten highly relational CIRR queries into concise, absolute descriptions boosts BASIC from 18.0% → 44.4% mAP and from 10% → 40% R@1 (Table 2 in that response). This confirms that the low scores stem largely from the relational wording prevalent in those datasets rather than from a fundamental weakness in composition.
3. Following the reviewer’s suggestion, we evaluate BASIC and BASIC without Harris regularization (BASIC*) on the newly released CoLLM-Refined versions of CIRR and Fashion-IQ (second Table below) . Both variants show notable performance gains on the refined datasets (up to 13.7% R@1 on refined-Fashion-IQ Shirt). However, it’s important to highlight that the CoLLM refinements primarily aim to reduce annotation ambiguities in CIRR and Fashion-IQ—this addresses a different challenge than modality dominance (evidenced by BASIC* outperforming BASIC) or CLIP’s sensitivity to specific types of phrasing, such as instructional or relational queries.

Speed and Storage Footprint

We consider Text × Image (“Product”) as the minimal CIR baseline, requiring only dot-products between the CLIP-encoded queries and the precomputed image features. BASIC builds on this with lightweight components—all applied at query time only.

Three components introduce non-negligible overhead:

• Semantic Projection:
  Requires storing a d × k projection matrix and applying it to the query, with both time and storage complexity of O(dk).

• Query Expansion:
  In our current implementation, this involves a secondary retrieval over top-Q items (O(QN)), along with a projection step of O(dk). This step can be made efficient using the FAISS library.

• Contextualization:
  This is the most computationally expensive component, requiring one extra (batched) forward pass through the CLIP text encoder, followed by averaging.

Other components—centering, min-based normalization, and Harris fusion—have trivial cost (O(d) or less) and negligible storage overhead.

Total runtime on an NVIDIA A100 40 GB GPU:

As shown, all components except contextualization add negligible latency, with total runtime increasing only from 25.5 ms to 25.9 ms. Contextualization, which involves CLIP forward passes, is the dominant factor—accounting for most of the 61.9 ms cost in the full BASIC pipeline (as shown also in Supplementary Figure S4).
To mitigate this, a key future direction is to develop a lightweight embedding network on top of CLIP that approximates contextualized representations without runtime overhead.

We thank Reviewer przL for the review and for explicitly recognising several key strengths of our work:
(i) the paper is well organised and explains the full benchmark-building pipeline,
(ii) thorough experiments on strong baselines demonstrate the effectiveness of BASIC on CIR task,
(iii) i-CIR is clearly more challenging than existing CIR datasets and will “contribute to further development” in CIR task, and
(iv) a training-free framework that nevertheless beats the SOTA is practically valuable.

Many of the points that the reviewer could not verify are clarified in the supplementary material (available via the ZIP link on the OpenReview page). We understand that its location may not have been immediately obvious and kindly invite the reviewer to consult it. Below, we provide point-by-point responses to all remaining comments.

Code Release

For NeurIPS main-track papers it is permitted to release code “upon acceptance”. About the public release policy, full data and code will be released under MIT license immediately upon acceptance. We wish to maximize the dissemination and usage of our dataset and method, thus we will also create a project page with engaging visualizations and demos.

Supplementray Material is Not Provided / Wrong Statement → It Is Provided

The full supplementary material—containing additional figures, tables, implementation details, and further discussion—is available via the ZIP link on the OpenReview page. We kindly invite the reviewer to consult it, as it addresses the points noted on lines 164, 244, 307, 330, 340, and others.

Typos & Minor Edits

We will fix in the camera-ready version of the paper.

Human Annotation vs. Scalability

High-quality, instance-level composed image retrieval ``labels'' cannot be obtained purely automatically, especially given (i) the exhaustive and explicit collection and curation of visual, textual, and composed hard negatives, and (ii) the copyright-safe and sentitive-content-safe approach we decided to follow. In instance-level image-to-image retrieval, datasets like GLDv2 (Google Landmarks Dataset v2 -- A Large-Scale Benchmark for Instance-Level Recognition and Retrieval, CVPR 2020) that are created automatically with crowd-sourced labels are known to contain a high amount of noise. Additionally, composed image retrieval has an extra complexity in creating a dataset. In CIR, alternative shortcuts in CIRR and FashionIQ result in well-known weaknesses, which are explicitly mentioned in our paper (subsection 3.4). These exact weaknesses are known in the community and have recently resulted (CoLLM: A Large Language Model for Composed Image Retrieval, CVPR 2025) in the introduction of a refined version of them (refined-CIRR, refined-FashionIQ), as mentioned by Reviewer rzCM. We perform experiments on these refined datasets as an answer to Reviewer rzCM and observe a performance boost of up to 13.7% R@10 on Fashion-IQ Shirt for our method. Our semi-automatic pipeline involving expert annotators yields a compact set that is as hard as adding 20M random LAION distractors (supplementary Fig. S5). As a compact benchmark, i-CIR lowers compute barriers and encourages fast iteration, but without sacrificing difficulty. For instance, extracting CLIP ViT-L features for the 300K-image i-CIR takes merely 1 h 25 min on a single NVIDIA V100 (32 GB), whereas processing a 20M-image distractor set would extend to almost four days. Thus, scaling up is not hindered. We believe that this trade-off is preferable for future benchmarks.

Category Hierarchy

Supp. Fig. S1 shows our 3-level taxonomy. “Same category” in Line 122 refers to level-3 (the most specific) categories.

Why a Visual Hard-Negative Can Become an Image Query / All Composed Positives Have Already Picked Up from Hard Negative Pools

The composed positives are picked from the overall candidate image pool (L115-116), which consists of potential queries, positives and (hard) negatives. There are three types of hard negatives: visual, textual, and composed (L130-132). Visual hard negatives include two cases: (i) the same instance under the same or different conditions; (ii) a very similar but different instance. Images of type (i) make excellent alternative image queries. Annotators therefore pick a subset of these and drop all original seed images.

Corpus Construction & GPT Usage

Section S2 of the supplementary material already describes the prompt we used with ChatGPT to generate the positive word corpus ($C_+$), which consists of object-centric terms used to guide semantic projection. However, due to an oversight, the corresponding prompt for generating the negative word corpus ($C_-$), which captures stylistic and contextual variations, was not included. We thank the reviewer for pointing this out. Below, we provide the exact prompt used:

We are working on a composed image retrieval task, where the input consists of an image and a text. The goal is to retrieve the most relevant matching image from a database. Typically, the image contains a main object in a particular “state”. This state can refer to a style (e.g., retro, handwritten, futuristic) or a contextual setting (e.g., next to the sea, on the beach, beside a table). For example: a retro mug, a cat next to a table, a handwritten notebook. We would like you to generate a .txt file containing 100 possible states in which an object can be found. Each line in the file should contain a single state expressed in natural language. Wait for me to say “next” before giving the next batch of states.

Here are the first 10 outputs of such a command:

on a wooden shelf
floating in water
wrapped in plastic
painted with graffiti
next to a fireplace
sitting on grass
surrounded by candles
hanging from a tree
under a spotlight

To support transparency and reproducibility, we will include this prompt in the revised supplementary material and provide the exact corpora used as part of the released codebase.

We thank Reviewer RdrS for the review and for explicitly recognizing strengths of our work:
(i) timely problem: CIR is an important problem,
(ii) clear visual presentation of our method, and
(iii) safeguards in the development of the dataset.
Below, we address each of their comments in detail.

Dataset Sample and Code Availability

Dataset. A sample of the dataset is already provided. Supp. Figs. S2 (one image query per instance) and S8 (all text queries) cover most query pairs in i-CIR; Figs. 1 & 3 show composed queries with positives and hard negatives.
Code. For NeurIPS main-track papers it is permitted to release code “upon acceptance”. Full data and code will be released under MIT license immediately upon acceptance. To maximize dissemination, we will also create a project page with engaging visualizations and demos.

Definition of CIR  “bidirectional retrieval”

Our opening phrase—“CIR combines image-to-image and text-to-image retrieval → (image and text) to image retrieval”—only means that the query is multimodal (image + text) while the database is visual. No “bidirectional” retrieval is implied. We can rephrase in the camera-ready for clarity.

Why Existing CIR Datasets Can Be Noisy

The construction protocols of CIRR, CIRCO, and FashionIQ rely on automatic selection of visually or semantically similar image pairs, followed by the annotation of a textual description that highlights their differences. The poor quality of CIRR and FashionIQ is well known and has led to refined-CIRR/FashionIQ (CoLLM, CVPR 2025), as mentioned by Reviewer rzCM. We refer explicitly to the construction details provided in the original papers:

• CIRR (§4.1): ``We first form image pairs then collect related annotations by crowd-sourcing. The pairs are drawn from subsets of images, as described below... We use the popular NLVR2 dataset... Prior to forming reference-target image pairs, we construct multiple subsets of six images that are semantically and visually similar, denoted as S... Here, to construct a subset, we randomly pick one image from the large corpus. We then sort the remaining images in dataset by their cosine similarity to this image using ResNet152 image feature vectors pre-trained on ImageNet... Within each constructed image subset S, we draw nine pairs of images... We collect a modification sentence for each pair of reference-target images using Amazon Mechanical Turk (AMT).''.
• CIRCO (§4): ``...We introduce an open-domain benchmarking dataset named Composed Image Retrieval on Common Objects in context (CIRCO). It is based on real-world images from COCO 2017 unlabeled set and is the first dataset for CIR with multiple ground truths. Contrary to CIRR, we start from a single pair of visually similar images and write a relative caption... Then, we propose to employ our approach to retrieve the top 100 images according to the query and combine them with the top 50 images most visually similar to the target one. Finally, we select the images that are valid matches for the query.''.
• FashionIQ (§3): "The images of fashion products that comprise Fashion IQ were originally sourced from a product review dataset... For each image, we followed the link to the product website available in the dataset, in order to extract corresponding product information, when available... To ensure that the relative captions can describe the fine-grained visual differences between the reference and target image, we leveraged product title information to select similar images for annotation with relative captions.".

This methodology often yields:
(i) pairs lacking meaningful fine-grained differences (supp. Fig. S10, rows 4–5);
(ii) text-only queries sufficing to find the target (supp. Fig. S9, row 4; quantitative in rebuttal Table in the answer to Reviewer rzCM);
(iii) negative pools containing false negatives (supp. Fig. S9, row 6).

Our semi-automatic pipeline with expert annotators yields a compact 300 K-image benchmark that is as hard as adding 20M random LAION distractors (supp. Fig. S5).

Ablation Table Structure & Harris Regularization

Table 2 is cumulative by design: we switch BASIC’s components on to illustrate how performance builds up through a logical sequence of design steps. Ordering is not arbitrary: Some components depend on the presence of others to be effective (e.g. projection assumes centred features, Harris requires min-normalised scores). At the reviewer's request, we also report a leave-one-out ablation (Table below), but note the expected caveats: removing min-norm or centering while leaving dependent steps active naturally hurts results. We hope this expanded analysis provides a clearer understanding of each component’s role and justifies the original cumulative layout of Table 2. Finally, Harris is a novel contribution to CIR that brings consistent improvements over all cases, even if its gains may appear more modest compared to other components. Its strength is aligned with the goal of CIR, as it reinforces retrievals that satisfy both the image and the text query, rather than favouring one modality over the other.

Motivation for Instance-level CIR

We kindly disagree with the reviewer. Even though the example that the reviewer provides, mentions the Eiffel tower → is an instance-level one, existing datasets follow a semantic-level definition (we cannot guarantee that a tiny fraction of instance-level queries does not exist, since we have not manually inspected the whole datasets). Examples from original papers:
(i) CIRCO [2] Fig. 7 retrieves different lunch box, clock, horse, etc.;
(ii) CIRR [23] Figs. 1&4 retrieve different dog or carriage;
(iii) FashionIQ [35] Fig. 3 retrieves a different T-shirt/blouse;
(iv) ImageNet-R (Pic2Word) [30] Fig. 3 retrieves any fish/tiger/shark matching the style cue.
Instance-level CIR therefore fills a genuine gap.

“Model-release safeguards”

We release no new model weights; BASIC is training-free (regarding composed retrieval) and it uses public OpenCLIP/SigLIP checkpoints. Only the original licences apply.

Baselines

• WeiCom = convex combination of normalised image-to-image and text-to-image similarities (λ ∈ [0, 1]). We shall add.
• TIRG requires supervised triplet training, so it is not a zero-shot baseline.

Purpose of BASIC

BASIC is purely a retrieval method; it is not used in dataset construction.

“Compact Yet Hard” Dataset

Thanks to curated visual, textual, and composed hard negatives, a 300K-image i-CIR database matches the difficulty of 20M LAION distractors. Extracting CLIP ViT-L features: 1 h 25 m on a single V100 vs. ≈ 4 days for 20 M images. Therefore experimentation and benchmarking is significantly sped up while keeping the difficulty of the dataset high.

Limitations

Methodology-wise, we already mention two limitations: modality-domination and cases where text query takes the form of an instructional or relational phrasing (in which training-free methods can underperform). Both are further explored in this rebuttal; see responses to Reviewer rzCM and Reviewer pyz4, respectively. Throughout the manuscript we also mention other limitations, such as the computational overhead of contextualization and the construction process of corpora via ChatGPT. Dataset-wise, our semi-automatic pipeline trades some speed for fidelity: every composed query is manually vetted (e.g. for hard negatives, licence/safety issues, etc.), so building i-CIR is slower and costlier than fully automatic datasets but yields an ambiguity-free benchmark the field currently lacks (Reviewer rzCM). Moreover, the design mostly assumes a single salient object per image query and uses English-only phrasing with public-domain imagery. We shall add.

Experimental Details

• Determinism: BASIC has no trainable parameters; single runs suffice.
• Compute: see supp. Fig. S4 for exact timings.
• Checkpoints/licences: OpenCLIP ViT-L/14 and SigLIP ViT-L/16 (Apache-2.0). BLIP not used—checklist typo.

Crowdsourcing & IRB

We appreciate the reviewer’s concern regarding ethical practices involving human annotators. To clarify, the human annotators in our study were employed (salaried institution employees) solely for the task of labelling and curating images for a new instance-level composed image retrieval dataset. The annotators were trained workers performing a well-defined annotation task. Our research is not performed on human subjects. No personal or sensitive data, such as images of people, faces, license plates, or other identifiable features, were included in the dataset. As such, our use of annotators falls under standard data labelling practices and does not meet the criteria for human subjects research that would require IRB oversight. We will update the paper and the checklist to clarify these points and avoid any misunderstanding regarding ethical compliance.