(CVPR 2026) ConeSep: Cone-based Robust Noise-Unlearning Compositional Network for Composed Image Retrieval

Zixu Li¹, Yupeng Hu^1✉, Zhiwei Chen¹, Mingyu Zhang¹, Zhiheng Fu¹, Liqiang Nie²

¹School of Software, Shandong University    
²School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen),    
^✉ Corresponding author  

Official Repository: A Cone-based robust noise-unlearning network addressing the Noisy Triplet Correspondence (NTC) problem in Composed Image Retrieval (CIR).

ConeSep (Cone-based robust noisE-unlearning comPositional network) is our proposed robust learning framework for Composed Image Retrieval. Based on an in-depth analysis of the "Noisy Triplet Correspondence (NTC)" problem, ConeSep systematically resolves three critical challenges overlooked by existing methods: Modality Suppression, Negative Anchor Deficiency, and Unlearning Backlash. By utilizing geometric boundary estimation and optimal transport, ConeSep actively perceives, structurally models, and precisely "unlearns" noise.

⬆ Back to top

• [2026-04-03] 🚀 Release all codes of ConeSep!

• [2026-02-21] 🔥 The paper "ConeSep: Cone-based Robust Noise-Unlearning Compositional Network for Composed Image Retrieval" has been accepted by CVPR 2026!.

⬆ Back to top

• 📐 Geometric Fidelity Quantization (GFQ): Theoretically establishes and practically estimates a noise boundary utilizing the geometric separability of the cone space, quantifying sample fidelity to overcome "Modality Suppression".
• 🛑 Negative Boundary Learning (NBL): Resolves the "Negative Anchor Deficiency" by learning a "diagonal negative combination" for each query, serving as an explicit semantic opposite-anchor in the embedding space.
• 🎯 Boundary-based Targeted Unlearning (BTU): Models the noisy correction process as an optimal transport (OT) problem to elegantly execute precise unlearning while completely avoiding "Unlearning Backlash" on clean samples.
• 🛡️ Unmatched NTC Robustness: Consistently surpasses State-of-the-Art (SOTA) robust CIR models (such as TME, HABIT, and INTENT) under severe Noise Triplet Correspondence (NTC) ratios (0%, 20%, 50%, 80%).

⬆ Back to top

⬆ Back to top

💡 Note for Fully-Supervised CIR Benchmarking:
🎯 The 0% noise setting in the tables below is equivalent to the traditional fully-supervised CIR paradigm. We highlight this 0% block to facilitate direct and fair comparisons for researchers working on conventional supervised methods.

⬆ Back to top

• Introduction
• News
• Key Features
• Architecture
• Experiment Results
• Install
• Data Preparation
• Quick Start
  • Training under Noisy Settings
  • Testing
• Project Structure
• Acknowledgement
• Contact
• Citation
• Support & Contributing

1. Clone the repository

git clone https://github.com/Lee-zixu/ConeSep
cd ConeSep

2. Setup Python Environment

The code is evaluated on Python 3.8.10 and CUDA 12.6. We recommend using Anaconda to create an isolated virtual environment:

conda create -n conesep python=3.8
conda activate conesep

# Install PyTorch (The evaluated environment uses Torch 2.1.0 with CUDA 12.1 compatibility)
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url [https://download.pytorch.org/whl/cu121](https://download.pytorch.org/whl/cu121)

# Install core dependencies
pip install scikit-learn==1.3.2 transformers==4.25.0 salesforce-lavis==1.0.2 timm==0.9.16

Note: Key dependencies include salesforce-lavis for the base architecture (BLIP-2).

⬆ Back to top

We evaluated our framework on two standard datasets: FashionIQ and CIRR. Please download the datasets first.

├── FashionIQ
│   ├── captions
│   │   ├── cap.dress.[train | val | test].json
│   │   ├── cap.toptee.[train | val | test].json
│   │   ├── cap.shirt.[train | val | test].json
│   ├── image_splits
│   │   ├── split.dress.[train | val | test].json
│   │   ├── split.toptee.[train | val | test].json
│   │   ├── split.shirt.[train | val | test].json
│   ├── dress
│   │   ├── [B000ALGQSY.jpg | B000AY2892.jpg | B000AYI3L4.jpg |...]
│   ├── shirt
│   │   ├── [B00006M009.jpg | B00006M00B.jpg | B00006M6IH.jpg | ...]
│   ├── toptee
│   │   ├── [B0000DZQD6.jpg | B000A33FTU.jpg | B000AS2OVA.jpg | ...]

├── CIRR
│   ├── train
│   │   ├── [0 | 1 | 2 | ...]
│   │   │   ├── [train-10108-0-img0.png | train-10108-0-img1.png | ...]
│   ├── dev
│   │   ├── [dev-0-0-img0.png | dev-0-0-img1.png | ...]
│   ├── test1
│   │   ├── [test1-0-0-img0.png | test1-0-0-img1.png | ...]
│   ├── cirr
│   ├── captions
│   │   ├── cap.rc2.[train | val | test1].json
│   ├── image_splits
│   │   ├── split.rc2.[train | val | test1].json

⬆ Back to top

In our implementation, we introduce the noise_ratio parameter to simulate varying degrees of NTC (Noise Triplet Correspondence) interference. You can reproduce the experimental results from the paper by modifying the --noise_ratio parameter (default options evaluated are 0.0, 0.2, 0.5, 0.8).

Training on FashionIQ:

python train.py \
    --dataset fashioniq \
    --fashioniq_path "/path/to/FashionIQ/" \
    --model_dir "./checkpoints/fashioniq_noise0.2" \
    --noise_ratio 0.2 \
    --batch_size 256 \
    --num_epochs 20 \
    --lr 2e-5

Training on CIRR:

python train.py \
    --dataset cirr \
    --cirr_path "/path/to/CIRR/" \
    --model_dir "./checkpoints/cirr_noise0.5" \
    --noise_ratio 0.5 \
    --batch_size 256 \
    --num_epochs 20 \
    --lr 1e-5

💡 Tips: > - Our model is based on the powerful BLIP-2 architecture. It is highly recommended to run the training on GPUs with sufficient memory (e.g., NVIDIA A40 48G / V100 32G).

• The best model weights and evaluation metrics generated during training will be automatically saved in the best_model.pt and metrics_best.json files within your specified --model_dir.

To generate the prediction files on the CIRR dataset for submission to the CIRR Evaluation Server, run the following command:

python src/cirr_test_submission.py checkpoints/cirr_noise0.5/

(The corresponding script will automatically output .json based on the generated best checkpoints in the folder for online evaluation.)

⬆ Back to top

Our code is deeply customized based on the LAVIS framework. The core implementations are centralized in the following files:

ConeSep/
├── lavis/
│   ├── models/
│   │   └── blip2_models/
│   │       └── ConeSep.py    # 🧠 Core model implementation: Includes GFQ, NBL, BTU modules
├── train.py                  # 🚀 Training entry point: Controls noise_ratio injection and optimal transport
├── datasets.py 
├── test.py 
├── utils.py 
├── data_utils.py 
├── cirr_test_submission.py   # Auxiliary scripts
├── datasets/                 # Dataset loading and processing logic
└── README.md

The implementation of this project references the LAVIS framework and the noise setting concepts from TME. We express our sincere gratitude to these open-source contributions!

⬆ Back to top

For any questions, issues, or feedback, please open an issue on GitHub or reach out to us at lizixu.cs@gmail.com

⬆ Back to top

Ecosystem & Other Works from our Team

TEMA (ACL'26) Paper | Web | Code	Air-Know (CVPR'26) Paper | Web | Code	HABIT (AAAI'26) Paper | Web | Code
ReTrack (AAAI'26) Paper | Web | Code	INTENT (AAAI'26) Paper | Web | Code	HUD (ACM MM'25) Paper | Web | Code
OFFSET (ACM MM'25) Paper | Web | Code	ENCODER (AAAI'25) Paper | Web | Code

If you find our work or this code useful in your research, please consider leaving a Star⭐️ or Citing📝 our paper 🥰. Your support is our greatest motivation!

@InProceedings{ConeSep,
    title={ConeSep: Cone-based Robust Noise-Unlearning Compositional Network for Composed Image Retrieval},
    author={Li, Zixu and Hu, Yupeng and Chen, Zhiwei and Zhang, Mingyu and Fu, Zhiheng and Nie, Liqiang},
    booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    year = {2026}
}

⬆ Back to top

We welcome all forms of contributions! If you have any questions, ideas, or find a bug, please feel free to:

• Open an Issue for discussions or bug reports.
• Submit a Pull Request to improve the codebase.

⬆ Back to top