Research map / Image generation / Autoregressive and latent generation
QuRe: Query-Relevant Retrieval through Hard Negative Sampling in Composed Image Retrieval
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QuRe, a composed image retrieval method that trains BLIP-2 with a pairwise reward-model (Bradley-Terry) loss against hard negatives sampled between steep relevance-score drops after the target, and introduces the HP-FashionIQ human-preference dataset.
Problem
Composed image retrieval (CIR) training with contrastive loss treats all other batch images as negatives, including false negatives, which yields irrelevant results even when the target is retrieved; existing metrics like Recall@k ignore the relevance of other retrieved images.
Approach
Replace in-batch contrastive loss with a pairwise reward-model objective (target vs. one sampled negative), and pick hard negatives from the band between the two largest relevance-score drops after the target in a periodically re-sorted corpus ranking.
Key results
- On FashionIQ validation, QuRe reaches 63.04 average recall vs. 62.13 for SPRC (second best).
- On CIRR test, QuRe achieves R@1 52.22 and R@5 82.53, beating SPRC by 1.47 and 1.95 points.
- On HP-FashionIQ, QuRe obtains a 74.55% preference rate, the highest among compared CIR models, and reaches 23.22 mAP@5 zero-shot on CIRCO versus 17.57 for SPRC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QuRe instead of SPRC | CIR retrieval · FashionIQ validation | Avg.: 62.13 -> 63.04 | Helped |
| Use QuRe instead of SPRC | CIR retrieval · CIRR test | Recall@1: 50.75 -> 52.22 | Helped |
| Use QuRe instead of SPRC | CIR retrieval · CIRR test | Recallsubset@1: 79.57 -> 78.51 | Hurt |
| Use QuRe instead of CLIP4CIR | CIR model ranking · HP-FashionIQ | Preference rate (%): 74.45 -> 74.55 | Helped |
| Use QuRe instead of CoVR-BLIP | CIR retrieval · CIRCO zero-shot (CIRR-pretrained) | mAP@5: 18.35 -> 23.22 | Helped |
| Use QuRe rather than Bi-BLIP4CIR with the same BLIP backbone | BLIP-based CIR · CIRR test | Recall@1: 32.55 -> 51.52 | Helped |
| Use QuRe rather than CoVR-2 with the same BLIP-2 backbone | BLIP-2-based CIR · CIRR test | Recall@1: 42.80 -> 52.22 | Helped |
| Select negatives from the entire image corpus (ablation) | QuRe hard-negative sampling · FashionIQ validation | average recall | Hurt |
| Select the top-k images by relevance score, without excluding false negatives (ablation) | QuRe hard-negative sampling · FashionIQ validation | average recall | Mixed |
| Select the top-k images after the target (ablation) | QuRe hard-negative sampling · FashionIQ validation | average recall | Mixed |
| Select negatives between two sharp relevance-score drops after the target (ablation) | QuRe hard-negative sampling · FashionIQ validation | average recall | Helped |
Surprising findings
- Models with higher Recall@k on FashionIQ (e.g., SPRC, CoVR-BLIP) do not always align better with human preference; CLIP4CIR aligns better with humans despite lower recall.
- Recall_subset@k is only second-best for QuRe on CIRR, since its hard negative definition can leave some false negatives ranked above the target.
Limitations
- Hard negative set size is not explicitly controlled and depends on score-drop detection; relies on the model's own scores which can include false negatives early in training.
- Evaluation of human preference is limited to FashionIQ shirts and top tees categories with 61 participants and a single preference-rate metric.
- Compared baselines use inconsistent backbones, which the authors partly address in an appendix.
Experimental setup
Datasets: FashionIQ CIRR CIRCO HP-FashionIQ
Benchmarks: FashionIQ CIRR HP-FashionIQ CIRCO
Metrics: Recall@10 Recall@50 Recall@K Recall_subset@K preference rate mAP@k
Baselines: CoSMo MGUR CLIP4CIR Bi-BLIP4CIR CoVR-BLIP SPRC CoVR-2
Builds on: BLIP-2 Bradley-Terry model HCL hard contrastive learning CLIP4CIR SPRC
Closest papers by meaning
- TriSampler: A Better Negative Sampling Principle for Dense Retrieval (AAAI 2024 · similarity 0.87)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.87)
- Aligning Composed Query with Image via Discriminative Perception from Negative Correspondences (AAAI 2025 · similarity 0.86)
- ConText-CIR: Learning from Concepts in Text for Composed Image Retrieval (CVPR 2025 · similarity 0.86)
- Learning with Noisy Triplet Correspondence for Composed Image Retrieval (CVPR 2025 · similarity 0.86)
- Generative Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.