QuRe: Query-Relevant Retrieval through Hard Negative Sampling in Composed Image Retrieval [ICML 2025]

Official implementation of QuRe: Query-Relevant Retrieval through Hard Negative Sampling in Composed Image Retrieval (ICML 2025).
[Paper Link]

Python Environment

The following commands set up a local Anaconda environment and install the required packages:

conda env create -f environment.yml -n qure

Prepare Datasets

Before running the code, please download the following datasets:

• FashionIQ
• CIRR
• CIRCO
• HP-FashionIQ

Once downloaded, update the base_path variable in each corresponding file with the local path to the dataset:

• ./data/fashionIQ.py
• ./data/cirr.py
• ./data/circo.py

For example:

base_path = '/path/to/dataset'

Training

To train the model on FashionIQ and CIRR datasets, use the following commands:

For FashionIQ:

python train_qure.py --config_path=configs/fashionIQ/train.json

For CIRR:

python train_qure.py --config_path=configs/cirr/train.json

Evaluation

To test the model on FashionIQ, CIRR, and CIRCO datasets, use the following commands:

For FashionIQ:

python evaluate_qure/evaluate_fiq.py --config_path=configs/fashionIQ/eval.json

For CIRR:

python evaluate_qure/evaluate_cirr.py --config_path=configs/cirr/eval.json

For CIRCO:

python evaluate_qure/evaluate_circo.py --config_path=configs/circo/eval.json

Checkpoints

We provide pre-trained checkpoints for both the FashionIQ and CIRR datasets.
You can download them from the link.

Acknowledgment

This code is built on top of the CoSMo and utilizes LAVIS. We thank the authors for their valuable contribution.

Citation

@inproceedings{kwakqure,
  title={QuRe: Query-Relevant Retrieval through Hard Negative Sampling in Composed Image Retrieval},
  author={Kwak, Jaehyun and Inhar, Ramahdani Muhammad Izaaz and Yun, Se-Young and Lee, Sung-Ju},
  booktitle={Forty-second International Conference on Machine Learning}
}