acceptodds
Under review as a conference paper at ICLR 2027

UniCIR: Unlocking Universal Embeddings for Training-Free Composed Image Retrieval

Abstract

Composed image retrieval (CIR) searches for images that satisfy a textual modification while preserving relevant content from a reference image. Many training-free methods construct queries through generated descriptions, making retrieval depend on the visual details captured in text. Universal multimodal embeddings jointly encode the reference image and modification, providing a strong alternative. However, retrieval can remain visually close to the reference while missing a requested change. We propose UniCIR, a training-free method that refines this joint representation through a controlled description pair. A frozen vision-language generator describes the intended target. We then replace one edit-linked value to form a negative description, keeping all other wording fixed. The shared retriever encodes both descriptions and refines the original query by adding the weighted positive and subtracting the weighted negative. This contrast targets the requested distinction while preserving the original joint embedding as an anchor. The resulting vector searches a fixed gallery with all model weights frozen. Our experiments show gains of 6.31, 7.89 and 8.68 percentage points over Qwen-8B on CIRR test R@1, CIRCO test mAP@5 and standard FashionIQ validation average R@10, respectively. Gains extend to four additional retrievers, while matched ablations show improvements over augmentation with positive descriptions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Training-Free Composed Image Retrieval, Universal Multimodal Embeddings, Contrastive Query Refinement, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_d653OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a very simple training-free method for composed image retrieval. The method works in a unified multimodal embedding space (main results in Qwen3-VL-Embedding but shown for more). The proposed method corrects the joint image-text embedding by using the text-only embedding as a positive and a hard negative text embedding as a negative. The corrected composed query is a simple weighted average of the three (negative weight for the negative). The method is very simple. Simplicity is not necessarily a negative quality, but this method is borderline simplistic.

Critical strengths
  1. The paper is well written and polished. I managed to go through it without a problem and understood everything I needed from the first linear read, something I find rare at this stage.
  2. The method definitely improves the baseline composed query (simple joint image-text embedding)
  3. The usefulness of the components is shown in an ablation study
Critical weaknesses
  1. The method is simple to the extent that it makes me wonder whether this is enough of a contribution for a standalone paper
  2. In Tables 1 and 2, if we exclude the supervised methods that are not comparable at all (think about removing them), the baseline (Qwen-8B) is shown to be state-of-the-art already (in CIRR and FashionIQ). Then you improve this baseline, but the main job is done by the Qwen3-VL-Embedding. That makes the methods not comparable at all
Questions for authors

I do not think Weakness 1 can be addressed at this stage, and for that, I will need to see what the other reviewers think. But for Weakness 2, I will have to ask you to provide some valid comparisons with 2-3 new (2025-2026) state-of-the-art methods. That is, training-free methods that use the same backbone models as your method. Either show UniCIR numbers with the competitors' backbones, or show competitor methods with Qwen3-VL-Embedding.

Reviewer_gbPaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a new training-free method for composed image retrieval (CIR). In CIR, the system receives a reference image and a text describing a modification, and it must find the target image in a gallery that reflects that modification. The authors introduce UniCIR, a query refinement method that improves the representations used by existing multimodal retrieval models. First, a vision-language model generates a positive description of the target image. The authors then construct a paired negative description by replacing a single edit-linked attribute value while keeping the remaining wording unchanged. Next, a shared retrieval model encodes the positive and negative descriptions separately. The original multimodal query embedding is adjusted toward the positive direction and away from the negative direction, which improves retrieval performance. Experiments show improvements over the corresponding universal retrieval baselines on CIRR, CIRCO, and FashionIQ. Ablation studies also measure how much the positive and negative descriptions each contribute to the final performance.

Critical strengths

Simple, training-free refinement with consistent empirical gains. The method is simple and requires no additional training, while delivering consistent improvements over the corresponding universal baselines across three benchmarks and five retrievers. Matched positive-only and two-positive controls further support the added value of the paired negative description beyond description augmentation alone. These results make the approach a practical and empirically well-supported refinement of frozen multimodal retrievers.

Critical weaknesses
  • Limited analysis of how semantic construction errors affect retrieval. The impact of semantic construction errors on retrieval performance remains inadequately characterized. Manual assessment of 180 accepted pairs revealed that only 65% satisfy all semantic validity criteria (Table 25). While the paper separately reports retrieval improvements and degradations, it does not correlate these outcomes with pair validity. Consequently, it is unclear whether semantically invalid pairs still contribute positively to refinement or disproportionately account for performance regressions. This ambiguity limits understanding of the method’s failure modes when the constructor fails to produce the intended contrast.

  • Limited coverage of edit types. The contrastive constructor relies on replacing a supported attribute value, leaving a substantial fraction of queries without refinement. Pair coverage is 67.88% on CIRR validation and 61.36% on CIRCO validation (Table 18), while the spatial/action groups marked as having no supported slot receive no refinement (Tables 21–22). Although these queries are retained in the evaluation through the Universal fallback, this limits the method’s applicability to edits beyond the supported attribute types.

Questions for authors
  • Only 65% of the sampled accepted contrastive pairs meet all the semantic validity criteria. The retrieval analysis also shows that UniCIR fixes some errors but introduces others. Could the authors report retrieval performance conditioned on semantic pair validity, for example comparing valid and invalid accepted pairs? In particular, do the retrieval degradations come mostly from invalid pairs? A small-scale analysis using the 180 manually assessed pairs would suffice.
  • Table 7 shows that independently generated negatives underperform the controlled construction overall. Could a hybrid strategy—retaining accepted controlled pairs and using less constrained generation only for fallback queries—improve coverage and retrieval performance? If feasible, a small-scale analysis on these fallback queries would help assess whether the coverage limitation can be alleviated.
  • The related work discusses DEO and MulVec, but neither appears in the experimental comparisons. Could the authors clarify whether comparable results are available under the evaluated protocols and, if so, include them? If direct comparison is not feasible, please explain the relevant differences in setting or resource requirements.
Reviewer_EF1GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the author proposed a training-free framework for composed image retrieval that refines universal multimodal embeddings using controlled positive–negative description pairs. A frozen vision–language model generates these pairs to capture local semantic contrasts, which are then used to refine the original image–text query embedding.

Critical strengths
  • The paper is clearly written and easy to understand.

  • The paper openly discusses the method’s limitations and failure cases.

  • The proposed UniCIR is intuitive, refining the original query by moving it toward the positive description and away from the negative description to emphasize the requested change.

Critical weaknesses
  • The methodological novelty is somewhat limited. Combining positive and negative evidence is already well established in composed image retrieval. Besides, Appendix C.3 shows that the proposed UniCIR’s refinement direction coincides with the normalized closed-form solution of a special case of DEO.

  • Methods such as DEO, DualCIR, and MulVec are discussed but are absent from the main quantitative comparisons.

  • The comparison is not entirely fair, as UniCIR uses a stronger universal backbone than many competing methods. Much of its apparent advantage may therefore come from the backbone itself. Would UniCIR still outperform these methods if they used the same strong backbone?

  • The paper reports inconsistent query times of 0.62 and 1.94 seconds. Without explaining this difference, the claimed 114× speedup over WISER is difficult to verify.

  • In section 4.4, the authors report that reference preservation decreases by 7.78, despite retaining the original query as an anchor. This suggests that the refinement can sacrifice relevant reference details to satisfy the requested change, without an explicit mechanism to prevent such errors.

Questions for authors

Could the authors clarify how pairs are accepted in Section 3.2 and how their semantic quality is evaluated in Appendix D.2? Are these two steps performed using predefined rules, human judgments, or an LLM?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.