Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image that reflects an intended textual modification while maintaining semantic consistency with a reference image. Existing training-free ZS-CIR methods typically rely on fine-grained generated target descriptions, which often overemphasize reference-specific visual details and lead to appearance-biased retrieval. Moreover, since these descriptions are generated through multimodal large language models (MLLMs), the direct influence of the original image-text query can be weakened in the final retrieval representation. In this paper, we propose DCR-TSF, a training-free ZS-CIR framework that combines Dual-level Compositional Reasoning (DCR) and Triadic Spherical Fusion (TSF). DCR generates fine- and coarse-grained descriptions from the reference image and modification text, capturing complementary target-side semantics across different granularities. To preserve the original image-text query information, TSF constructs a composed-query anchor by spherically interpolating the reference image feature and modification text feature, and fuses it with the fine- and coarse-grained representations using a weighted Karcher mean on the unit hypersphere. The resulting final query representation balances detailed visual evidence, high-level semantic abstraction, and direct preservation of the original composed query. Extensive experiments on CIRCO, CIRR, and Fashion-IQ demonstrate that our method achieves state-of-the-art performance among training-free methods.
Figures & tables
Figure 1: Biases in Fine- and Coarse-grained Descriptions for ZS-CIR. Colored text refers to observed elements in the reference image. Background color indicates phrases appearing in the modification text. The bar colors in the text-to-image similarity score plots correspond to the candidate images. (a): Fine-grained description is biased towards the visual details of the reference image. (b): Coarse-grained description lacks details of the reference image.
Figure 2: Overview of DCR-TSF. Given a reference image Ir and a modification text Dm , DCR generates fine- and coarse-grained descriptions, which are encoded into Tfine and Tcoarse . TSF constructs a composed-query anchor Qcomp from Vr and Tm , and fuses it with Tfine and Tcoarse using a weighted Karcher mean [ 11 ] . The resulting final query representation Tfinal is used to compute the final retrieval score S(i) .
Figure 3: Comparison between OSrCIR and our coarse-grained description prompt. OSrCIR [ 28 ] generates a detailed target description through a CoT prompt, preserving rich reference-specific visual evidence that may be irrelevant to the modification condition. In contrast, our coarse-grained description prompt identifies stable visual anchors and focuses on modification-related semantics, omitting excessive reference-specific details and providing a complementary abstraction for DCR.
Backbone
Method
Training -free
CIRCO
CIRR
mAP@5
mAP@10
mAP@25
mAP@50
R@1
R@5
R@10
R sub @1
R sub @2
R sub @3
Avg.
ViT-B/32
PALAVRA [ 6 ]
✗
4.61
5.32
6.33
6.80
16.62
43.49
58.51
41.61
65.30
80.94
42.55
SEARLE [ 4 ]
✗
9.35
9.94
11.13
11.84
24.00
53.42
66.82
54.89
76.60
88.19
54.16
Slerp [ 10 ]
✓
6.35
7.11
8.12
8.75
24.22
50.94
63.49
57.86
78.27
89.25
54.40
CIReVL [ 12 ]
✓
14.94
15.42
17.00
17.82
23.94
52.51
66.00
60.17
80.05
90.19
56.34
LDRE [ 34 ]
✓
17.96
18.32
20.21
21.11
25.69
55.13
69.04
60.53
80.65
90.70
57.83
Table 1: Comparison with the state-of-the-art methods on the CIRCO and CIRR test sets. Best results are highlighted in bold, and the second best are underlined.
Backbone
Method
Training -free
Shirt
Dress
Tops&Tee
Avg.
R@10
R@50
R@10
R@50
R@10
R@50
R@10
R@50
R mean
ViT-B/32
PALAVRA [ 6 ]
✗
21.49
37.05
17.25
35.94
20.55
38.76
19.76
37.25
28.50
SEARLE [ 4 ]
✗
24.44
41.61
18.54
39.51
25.70
46.46
22.89
42.53
32.71
Slerp [ 10 ]
✓
20.53
41.00
23.75
40.92
26.98
46.77
23.75
42.90
33.33
CIReVL [ 12 ]
✓
28.36
47.84
25.29
46.36
31.21
53.85
28.29
49.35
38.82
LDRE [ 34 ]
✓
27.38
46.27
19.97
41.84
27.07
48.78
24.81
45.63
35.22
Table 2: Comparison with the state-of-the-art methods on Fashion-IQ. R mean represents the average results across all the metrics. Best results are highlighted in bold, and the second best are underlined.
FD
CD
Comp
CIRCO
CIRR
Fashion-IQ
mAP@5
mAP@10
R@5
R sub @1
R@10
R@50
✓
25.55
26.66
60.00
61.54
35.57
56.03
✓
32.03
33.34
64.29
69.06
32.83
52.84
✓
17.32
17.88
62.75
67.42
40.58
60.39
✓
✓
34.77
35.95
67.04
68.15
38.43
59.35
✓
✓
32.01
32.88
69.33
67.66
43.86
65.23
Table 3: Ablation study with CLIP ViT-G/14. FD, CD, and Comp denote fine-grained descriptions, coarse-grained descriptions, and the composed-query anchor, respectively. The last row corresponds to the full DCR-TSF.
Figure 4: Hyperparameter sensitivity of TSF fusion weights. Ternary plots show validation results for different TSF fusion-weight configurations on each dataset using CLIP ViT-G/14. The star marker indicates the best-performing configuration selected for the main experiments.
Figure 8
Method
mAP@5
mAP@10
mAP@25
mAP@50
Qwen2.5-VL-3B [ 3 ]
23.72
24.89
27.09
28.09
Qwen2.5-VL-7B [ 3 ]
30.62
31.86
34.29
35.42
Qwen2.5-VL-32B [ 3 ]
34.77
35.95
38.56
39.79
Qwen2.5-VL-72B [ 3 ]
35.26
36.69
39.46
40.62
Qwen2-VL-72B [ 32 ]
34.38
36.11
38.81
39.89
Qwen3-VL-4B [ 31 ]
33.21
34.48
37.23
38.37
Table 5: Analysis on scalability and model variation in MLLMs on the CIRCO test set with CLIP ViT-G/14.
BLIP-2 & GPT-3.5-turbo
Multi-caption generator
LLM-based editing reasoner
Total
SEIZE [ 33 ]
1.560 ± 0.056
3.912 ± 0.812
5.472 ± 0.791
Qwen2.5-VL-32B
Image-scale reasoning
Object-scale reasoning
Total
CoTMR [ 27 ]
1.892 ± 0.011
1.661 ± 0.014
3.553 ± 0.023
Qwen2.5-VL-32B
Fine-grained description
Coarse-grained description
Total
DCR-TSF (Ours)
1.883 ± 0.019
1.584 ± 0.018
3.467 ± 0.030
Table 6: Inference time comparison with existing training-free ZS-CIR methods. We report the average latency per composed query in seconds. Each column decomposes the total inference time into the main reasoning stages of each method.
Figure 6: Qualitative comparison on CIRCO with ViT-G/14. Each example shows a reference image, its modification text, and the retrieved images with their ranks under four settings: Fine-Grained Description , Coarse-Grained Description , DCR , and DCR-TSF . The ground-truth image is highlighted in green.
Figure A : Prompt of fine descriptions Pfine
Figure B : Prompt of coarse descriptions Pcoarse
Backbone
FD
CD
Comp
CIRCO
CIRR
Fashion-IQ
mAP@5
mAP@10
R@5
R sub @1
R@10
R@50
ViT-B/32
✓
18.21
18.83
55.88
58.96
32.08
52.84
✓
21.28
21.89
59.98
65.88
29.91
50.87
✓
11.00
11.41
59.21
63.90
33.97
54.49
✓
✓
22.95
23.76
62.29
65.16
35.57
57.02
✓
✓
22.15
22.86
65.74
64.89
39.50
60.82
Table A : Extended ablation study across CLIP backbones. FD, CD, and Comp denote the fine-grained representation, coarse-grained representation, and composed-query anchor, respectively. We report results on CIRCO, CIRR, and Fashion-IQ using ViT-B/32, ViT-L/14, and ViT-G/14.
Backbone
Fusion Strategy
CIRCO
CIRR
Fashion-IQ
mAP@5
R@5
R mean
ViT-B/32
Weighted Avg.
25.11
67.18
51.32
Sequential Slerp
25.05
67.22
51.31
Karcher Mean
25.07
67.28
51.33
ViT-L/14
Weighted Avg.
31.92
69.95
51.46
Sequential Slerp
31.56
69.88
51.41
Table B : Comparison of fusion strategies across CLIP backbones. Results are reported on CIRCO, CIRR, and Fashion-IQ with ViT-B/32 and ViT-L/14.
Method
Reasoning Backbone
mAP@5
mAP@10
CIReVL [ 12 ]
BLIP-2 [ 16 ] + GPT-3.5-Turbo [ 5 ]
26.77
27.59
OSrCIR [ 28 ]
GPT-4o [ 1 ]
30.47
31.14
SEIZE [ 33 ]
BLIP-2 [ 16 ] + GPT-3.5-Turbo [ 5 ]
32.46
33.77
CoTMR [ 27 ]
Qwen2-VL-72B [ 32 ]
32.23
32.72
DCR (Ours)
Qwen2-VL-72B [ 32 ]
34.38
36.11
DCR (Ours)
Qwen2.5-VL-32B [ 3 ]
34.77
35.95
Table C : Comparison of reasoning backbones on the CIRCO test set with CLIP ViT-G/14.
Backbone
Fusion Weights
CIRCO
CIRR
Fashion-IQ
mAP@5
R@5
R mean
ViT-B/32
Equal Weights
25.02
66.77
50.81
Dataset-specific
25.07
67.28
51.33
ViT-L/14
Equal Weights
31.60
69.25
50.46
Dataset-specific
31.80
69.95
51.52
Table D : Sensitivity analysis of fusion weights. Comparison of equal and dataset-specific fusion weights on CIRCO, CIRR, and Fashion-IQ using CLIP ViT-B/32 and ViT-L/14.
CIRCO (mAP@5)
CIRR (R@5)
Fashion-IQ (R mean )
Performance ↑
35.35 / 34.63
72.61 / 70.80
55.71 / 58.30
Modality Gap ↓
0.5524 / 0.0084
0.5366 / 0.0133
0.5990 / 0.0195
Table E : Analysis of modality gap reduction. Comparison of retrieval performance and modality gap with and without I0T post on the validation splits of three benchmarks. Values are reported as DCR-TSF / I0T post .
Figure C : Comparison of DCR with caption ensemble baselines on CIRCO and Fashion-IQ.
Figure D : Qualitative comparison on Fashion-IQ with ViT-G/14. Each example shows a reference image, its modification text, and the retrieved images with their ranks under four settings: Fine-Grained Description , Coarse-Grained Description , Composed-query Anchor , and DCR-TSF . The ground-truth image is highlighted in green.
Figure E : Qualitative comparison on CIRR with ViT-G/14. Each example shows a reference image, its modification text, and the retrieved images with their ranks under four settings: Fine-Grained Description , Coarse-Grained Description , Composed-query Anchor , and DCR-TSF . The ground-truth image is highlighted in green.
Figure G : Failure cases caused by negation on CIRR with ViT-G/14. CLIP assigns high similarity to images containing objects or interactions explicitly excluded by negated modification texts (e.g., “no books” or “no human interacting”). Since TSF directly incorporates the modification text, this limitation propagates to retrieval and can rank such images above the ground-truth target.
Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image from a multimodal query consisting of a reference image and an edit text describing the desired modification. Recent ZS-CIR studies have relied on projection-based methods that map a reference image into pseudo-word tokens in the text embedding space. However, such methods require additional projection and re-encoding steps, increasing training complexity, reducing efficiency, and introducing a discrepancy between training and inference. In this paper, we propose DiCE-CIR, a direct composition learning method that predicts composed query representations by directly composing a reference image and an edit text. To enable scalable training without manually annotated triplets, we automatically construct compositional training samples from large-scale image-caption pairs using a large language model. Based on these samples, we train a lightweight composition module with objectives that promote alignment with the target, edit-consistent semantic transformation, and retrieval discriminability. We conduct extensive experiments on ZS-CIR benchmarks and show that DiCE-CIR achieves state-of-the-art performance on CIRCO and competitive performance on CIRR while maintaining high computational efficiency.
Gwang-Ho Na, Ho-Joong Kim, Seong-Whan Lee
Department of Artificial Intelligence, Korea University, Anam-dong, Seongbuk-ku, Seoul 02841, Korea
Zero-Shot Composed Image Retrieval (ZS-CIR) aims to retrieve target images given a multimodal query (comprising a reference image and a modification text), without training on annotated triplets. Existing methods typically convert the multimodal query into a single modality-either as an edited caption for Text-to-Image retrieval (T2I) or as an edited image for Image-to-Image retrieval (I2I). However, each paradigm has inherent limitations: T2I often loses fine-grained visual details, while I2I struggles with complex semantic modifications. To effectively leverage their complementary strengths under diverse query intents, we propose WISER, a training-free framework that unifies T2I and I2I via a "retrieve-verify-refine" pipeline, explicitly modeling intent awareness and uncertainty awareness. Specifically, WISER first performs Wider Search by generating both edited captions and images for parallel retrieval to broaden the candidate pool. Then, it conducts Adaptive Fusion with a verifier to assess retrieval confidence, triggering refinement for uncertain retrievals, and dynamically fusing the dual-path for reliable ones. For uncertain retrievals, WISER generates refinement suggestions through structured self-reflection to guide the next retrieval round toward Deeper Thinking. Extensive experiments demonstrate that WISER significantly outperforms previous methods across multiple benchmarks, achieving relative improvements of 45% on CIRCO (mAP@5) and 57% on CIRR (Recall@1) over existing training-free methods. Notably, it even surpasses many training-dependent methods, highlighting its superiority and generalization under diverse scenarios. Code will be released at https://github.com/Physicsmile/WISER.
Tianyue Wang, Leigang Qu, Tianyu Yang +4
Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences · School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences · National University of Singapore +3
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.