Less is More: Concise Retrieval and Critic-based Re-ranking for Composed Image Retrieval

Yuqian Zheng; Hyoju Ahn; Hyeongjin Ahn; Eunil Park

cs.CVOct 20, 2026

Less is More: Concise Retrieval and Critic-based Re-ranking for Composed Image Retrieval

Authors: Yuqian Zheng, Hyoju Ahn, Hyeongjin Ahn, Eunil Park

Organizations: Department of Applied Artificial Intelligence, Sungkyunkwan University, Seoul, Republic of Korea · Department of Immersive Media Engineering, Sungkyunkwan University, Seoul, Republic of Korea · Jaume I University, Castellon, Spain

Abstract

Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image based on a reference image and modification text. Recent approaches commonly rely on large vision-language models to generate target image captions for retrieval. However, these captions may include redundant semantic details, introducing noise when retrieving the target image and leading to suboptimal retrieval results. To address this issue, we propose Less-is-More CIR, a two-stage training-free framework that combines Concise Retrieval and Critic Re-ranking. In the first stage, concise target-oriented captions are generated that preserve only the most discriminative modification semantics, thereby reducing semantic noise in the retrieval. In the second stage, a critic re-rank module further evaluates the retrieved candidates from the perspectives of modification consistency and reference preservation. Extensive experiments demonstrate the effectiveness and robustness of the proposed framework for the ZS-CIR. Our code is available at: https://github.com/dxlabskku/LM-CIR.

Figures & tables

Figure 1. Comparison of target image captions generated by existing ZS-CIR methods and our concise reasoning strategy. Normal CoT introduces semantic noise, whereas concise reasoning generates cleaner target descriptions. Comparison of target image captions generated by existing ZS-CIR methods and our concise reasoning strategy. Normal CoT introduces semantic noise, whereas concise reasoning generates cleaner target descriptions.

Panel of Figure 2 — Figure 2. Overview of the two-stage LM-CIR framework. (a) In Stage 1, an LVLM generates a concise target image caption and retrieves $N$ candidates using fused image-image and text-image similarity scores. (b) In Stage 2, the Top-20 candidates are re-ranked based on EDIT and PRESERVE scores predicted by the LVLM critic. Overview of the two-stage LM-CIR framework. (a) In Stage 1, an LVLM generates a concise target image caption and retrieves $N$ candidates using fused image-image and text-image similarity scores. (b) In Stage 2, the Top-20 candidates are re-ranked based on EDIT and PRESERVE scores predicted by the LVLM critic.

Benchmark		CIRCO				CIRR
Metric		mAP@ $k$				Recall@ $k$				Recall sub @ $k$
Backbone	Method	$k$ =5	$k$ =10	$k$ =25	$k$ =50	$k$ =1	$k$ =5	$k$ =10	$k$ =50	$k$ =1	$k$ =2	$k$ =3
ViT-G/14	CIReVL ( Karthik et al., 2024 )	26.77	27.59	29.96	31.03	34.65	64.29	75.06	91.66	67.95	84.87	93.21
	OSrCIR ( Tang et al., 2025 )	30.47	31.14	35.03	36.59	37.26	67.25	77.33	–	69.22	85.28	93.55
	SEIZE ( Yang et al., 2024 )	32.46	33.77	36.46	37.55	38.87	69.42	79.42	–	74.15	89.23	95.71
	CoTMR ( Sun et al., 2025 )	32.23	32.72	35.60	36.83	36.36	67.52	77.82	93.99	71.19	86.34	93.87

Table 1. Performance comparison on the CIRCO and CIRR test benchmarks.

sub

denotes the CIRR subset retrieval task. Moreover, the best and second-best results are indicated by bold and underlined text, respectively.

Backbone	Method	Shirt		Dress		Toptee		Avg.
Backbone	Method	R@10	R@50	R@10	R@50	R@10	R@50	R@10	R@50
ViT-G/14	CIReVL ( Karthik et al., 2024 )	33.71	51.42	27.07	49.53	35.80	56.14	32.19	52.36
	OSrCIR ( Tang et al., 2025 )	38.65	54.71	33.02	54.78	41.04	61.83	37.57	57.11
	AutoCIR ( Cheng et al., 2025 )	36.36	55.84	26.18	47.69	37.28	60.38	33.27	54.63
	SEIZE ( Yang et al., 2024 )	43.60	65.42	39.61	61.02	45.94	71.12	43.05	65.85
	CoTMR ( Sun et al., 2025 )	38.32	62.24	34.51	57.36	41.90	64.30	38.25	61.32

Table 2. Performance comparison on the FashionIQ validation benchmark. The best and second-best results are indicated by bold and underlined text, respectively.

Figure 3. Examples of Concise Retrieval on the CIRR and FashionIQ validation benchmarks. Retrieval results are shown from left to right according to the ranking order, where the ground-truth target images are successfully retrieved at top ranks. Examples of Concise Retrieval on the CIRR and FashionIQ validation benchmarks. Retrieval results are shown from left to right according to the ranking order, where the ground-truth target images are successfully retrieved at top ranks.

Figure 4. Example of the proposed critic-based re-ranking. The ground-truth target image, which is ranked low in the Stage-1 retrieval results, is successfully promoted to a higher rank after critic-based re-ranking. This figure shows a qualitative example of critic-based re-ranking. A reference image and modification text are used to generate the target caption. Candidate images retrieved in Stage 1 are then evaluated by an LVLM critic using EDIT and PRESERVE scores. The ground-truth candidate is promoted from rank 4 to rank 1 after re-ranking the list.

1. Critic Re-ranking Number
Methods	mAP@5	mAP@10	mAP@25	mAP@50
Top-5	32.82	33.16	35.58	36.70
Top-10	36.82	35.47	37.86	38.98
Top-20 (Full model)	39.82	39.61	40.74	41.86
2. Critic Re-ranking LVLM
Qwen2.5-32B (Full model)	39.82	39.61	40.74	41.86

Table 3. Ablation studies on the CIRCO test set using the ViT-L/14 backbone.

Methods	mAP@5	mAP@10	mAP@25	mAP@50
Concise (GPT-5.1)	25.28	26.13	28.49	29.53
Concise (Qwen3)	24.28	25.47	27.54	28.47
CIReVL	18.57	19.01	20.89	21.80
OSrCIR	23.87	25.33	27.84	28.97
AutoCIR	24.05	25.14	27.35	28.36

Table 4. Comparison of different caption generation strategies on the CIRCO test set under pure text-image retrieval.

Figure 5. Failure cases of Concise Retrieval and Critic Re-ranking. Failure cases of Concise Retrieval and Critic Re-ranking.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6. Concise Chain-of-Thought Prompting. Concise Chain-of-Thought Prompting.

Figure 7. Prompt Examples for CIRCO and CIRR. Prompt Examples for CIRCO and CIRR.

Figure 8. Prompt Examples for FashionIQ. Prompt Examples for FashionIQ.

Figure 9. Critic Re-ranking Prompting. Critic Re-rank Prompting.

Composed Image Retrieval

Composed image retrieval finds images that match a reference image together with a text instruction describing the desired change.

See all 65 papers

G-MIXER: Geodesic Mixup-based Implicit Semantic Expansion and Explicit Semantic Re-ranking for Zero-Shot Composed Image Retrieval
Apr 16, 2026Jiyoung Lim, Heejae Yang, Jee-Hyong LeeZero-Shot Learning Composed Image Retrieval
Sungkyunkwan University
DiCE-CIR: Direct Composition Learning for Efficient Zero-Shot Composed Image Retrieval
Jul 6, 2026Gwang-Ho Na, Ho-Joong Kim, Seong-Whan LeeCross-Modal Alignment Compositional Representation Learning
Department of Artificial Intelligence, Korea University, Anam-dong, Seongbuk-ku, Seoul 02841, Korea
Thinking Before Retrieving: Robust Zero-Shot Composed Image Retrieval via Strategic Planning and Self-Criticism
Jun 30, 2026Gunho Jung, Jeong-Woo Park, Seon Bin Kim +1Image Retrieval Composed Image Retrieval
Department of Artificial Intelligence, Korea University, Seoul, 02841, Republic of Korea

Apr 16, 2026J/K move · Enter open · S save

G-MIXER: Geodesic Mixup-based Implicit Semantic Expansion and Explicit Semantic Re-ranking for Zero-Shot Composed Image Retrieval

Jiyoung Lim, Heejae Yang, Jee-Hyong Lee

Sungkyunkwan University

Zero-Shot Learning Composed Image Retrieval

Open paper AskPDF arXiv

Composed Image Retrieval (CIR) aims to retrieve target images by integrating a reference image with a corresponding modification text. CIR requires jointly considering the explicit semantics specified in the query and the implicit semantics embedded within its bi-modal composition. Recent training-free Zero-Shot CIR (ZS-CIR) methods leverage Multimodal Large Language Models (MLLMs) to generate detailed target descriptions, converting the implicit information into explicit textual expressions. However, these methods rely heavily on the textual modality and fail to capture the fuzzy retrieval nature that requires considering diverse combinations of candidates. This leads to reduced diversity and accuracy in retrieval results. To address this limitation, we propose a novel training-free method, Geodesic Mixup-based Implicit semantic eXpansion and Explicit semantic Re-ranking for ZS-CIR (G-MIXER). G-MIXER constructs composed query features that reflect the implicit semantics of reference image-text pairs through geodesic mixup over a range of mixup ratios, and builds a diverse candidate set. The generated candidates are then re-ranked using explicit semantics derived from MLLMs, improving both retrieval diversity and accuracy. Our proposed G-MIXER achieves state-of-the-art performance across multiple ZS-CIR benchmarks, effectively handling both implicit and explicit semantics without additional training. Our code will be available at https://github.com/maya0395/gmixer.

Abstract

Composed Image Retrieval

Explore similar work

G-MIXER: Geodesic Mixup-based Implicit Semantic Expansion and Explicit Semantic Re-ranking for Zero-Shot Composed Image Retrieval

DiCE-CIR: Direct Composition Learning for Efficient Zero-Shot Composed Image Retrieval

Thinking Before Retrieving: Robust Zero-Shot Composed Image Retrieval via Strategic Planning and Self-Criticism