FlowCIR: Semantic Transport via Flow Matching for Zero-Shot Composed Image Retrieval

ECCV 2026
Zhenqi He, Ziqi Jiang, Yuanpei Liu, Yanghao Wang, Teng Wang, Long Chen
LONG Group, HKUST

FlowCIR learns semantic transport with flow matching for zero-shot composed image retrieval.

Setup

conda create --name flowcir python=3.10 -y
conda activate flowcir
pip install torch torchvision numpy pillow tqdm datasets transformers
pip install git+https://github.com/openai/CLIP.git

Run the following data preparation, training, validation, and submission commands from the code directory:

cd code

Data Preparation

Download HQ-Edit and the target CIR datasets (CIRR, CIRCO, and Fashion-IQ). Configure the CIRR and CIRCO locations in config.py, or use environment variables:

export FLOWCIR_HQEDIT_DIR=/path/to/HQ-Edit
export FLOWCIR_CIRR_DIR=/path/to/CIRR
export FLOWCIR_CIRCO_DIR=/path/to/CIRCO

Before extracting test features, copy captions/cap.rc2.test1_posneg.json to <CIRR_DIR>/cirr/captions/ and captions/test_posneg.json to <CIRCO_DIR>/annotations/, replacing the corresponding non-posneg test annotations when necessary.

Use extract_features/extracted_features_clip_posneg.py to extract the positive and negative text features from the posneg annotations.

Feature Extraction

Extract HQ-Edit training features:

bash myScripts/extract_hqedit_feat.sh

Extract CIRR or CIRCO retrieval features:

DATASET=cirr bash myScripts/extract_cir_feat.sh
DATASET=circo bash myScripts/extract_cir_feat.sh

Training and Validation

bash myScripts/train.sh
MODEL_DIR=/path/to/checkpoint_directory \
  EPOCHS="10 15 20" END_POINTS="0.4 0.5 0.6" \
  bash myScripts/val.sh

Checkpoints and logs are written to the directories configured by FLOWCIR_SAVE_ROOT and FLOWCIR_LOG_DIR. Empirically, training for about 200 epochs is usually sufficient. Select the checkpoint epoch and inference end time according to validation performance; end times between 0.4 and 0.6 generally work well.

Submission

Generate the final submission with the epoch and end time selected on the validation set:

MODEL_DIR=/path/to/checkpoint_directory \
  DATASET=cirr EPOCH=15 END_POINT=0.5 \
  bash myScripts/submit.sh

Use DATASET=circo to generate a CIRCO submission. Submission files are saved under FLOWCIR_SUBMISSION_DIR.

Acknowledgement

This repository is developed in part based on Exploring Cross-Modal Flows for Few-Shot Learning (FMA) (ICLR 2026). It also uses OpenAI CLIP and the publicly available HQ-Edit, CIRR, and CIRCO resources. We thank their authors for making these projects available.

Citation

If you find this work useful, please cite flowcir.bib:

@inproceedings{he2026flowcir,
  title     = {FlowCIR: Semantic Transport via Flow Matching for Zero-Shot Composed Image Retrieval},
  author    = {He, Zhenqi and Jiang, Ziqi and Liu, Yuanpei and Wang, Yanghao and Wang, Teng and Chen, Long},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}