---
license: apache-2.0
tags:
- composed-image-retrieval
- vision-language
- multimodal
- pytorch
---

<a id="top"></a>
<div align="center">
  <h1>⚓ TEMA: Anchor the Image, Follow the Text for Multi-Modification Composed Image Retrieval</h1>

  <p>
    <b>Zixu Li</b><sup>1</sup>&nbsp;
    <b>Yupeng Hu</b><sup>1✉</sup>&nbsp;
    <b>Zhiheng Fu</b><sup>1</sup>&nbsp;
    <b>Zhiwei Chen</b><sup>1</sup>&nbsp;
    <b>Yongqi Li</b><sup>2</sup>&nbsp;
    <b>Liqiang Nie</b><sup>3</sup>
  </p>

  <p>
    <sup>1</sup>Shandong University&nbsp;
    <sup>2</sup>Hong Kong Polytechnic University&nbsp;&nbsp;
    <sup>3</sup>Harbin Institute of Technology (Shenzhen)
  </p>
</div>

These are the official model and data resources for **TEMA** (Text-oriented Entity Mapping Architecture), the first Composed Image Retrieval (CIR) framework designed explicitly for multi-modification scenarios. 

🔗 **Paper:** [Accepted by ACL 2026] 
🔗 **GitHub Repository:** [lee-zixu/ACL26-TEMA](https://github.com/lee-zixu/ACL26-TEMA)

---

## 📌 Model Information

### 1. Model Name
**TEMA** (Text-oriented Entity Mapping Architecture) Checkpoints.

### 2. Task Type & Applicable Tasks
- **Task Type:** Composed Image Retrieval (CIR) / Vision-Language / Multimodal Alignment
- **Applicable Tasks:** Retrieving target images based on a reference image and complex Multi-Modification Texts (MMT), seamlessly accommodating both simple and complex modifications.

### 3. Project Introduction
Prevailing CIR setups rely on simple modification texts, inducing two critical limitations in practical applications: **Insufficient Entity Coverage** and **Clause-Entity Misalignment**. 

**TEMA** brings CIR closer to real-world use cases by introducing:
- 🧠 **MMT Parsing Assistant (PA):** Utilizes an LLM-based text summarizer and a Consistency Detector during training to enhance the exposure and coverage of modified entities.
- 🔗 **MMT-oriented Entity Mapping (EM):** Introduces learnable queries to consolidate multiple clauses of the same entity on the text side and align them with corresponding visual entities on the image side.

### 4. Training Data Source
The model is evaluated and trained using our newly proposed instruction-rich multi-modification datasets: 
- **M-FashionIQ** (Fashion domain)
- **M-CIRR** (Open domain)

These datasets replace short, simplistic texts with Multi-Modification Texts (MMT) generated by MLLM and verified by human annotators.

---

## 🚀 Usage & Basic Inference

These weights and codes are designed to be used with the official TEMA GitHub repository.

### Step 1: Prepare the Environment
Clone the GitHub repository and install the required dependencies (evaluated on Python 3.10.8 and PyTorch 2.5.1):
```bash
git clone [https://github.com/lee-zixu/ACL26-TEMA](https://github.com/lee-zixu/ACL26-TEMA)
cd TEMA
conda create -n tema python=3.10.8 -y
conda activate tema

# Install PyTorch
pip install torch==2.5.1 torchvision torchaudio --index-url [https://download.pytorch.org/whl/cu121](https://download.pytorch.org/whl/cu121)

# Install core dependencies
pip install transformers==4.25.0
```

### Step 2: Download Model & Data
Please refer to the [GitHub Repository](https://github.com/lee-zixu/ACL26-TEMA) for detailed instructions on downloading the base image datasets (FashionIQ and CIRR) and replacing their captions with our provided `mmt_captions` to construct the **M-FashionIQ** and **M-CIRR** datasets. 

Ensure your folder structure matches the requirements in the official codebase.

### Step 3: Run Training / Inference
Once the environment and datasets are prepared, you can start the training or evaluation process:
```bash
python3 train.py
```

---

## ⚠️ Limitations & Notes

**Disclaimer:** This framework and the constructed M-FashionIQ/M-CIRR datasets are intended for **academic research and multimodal evaluation**. 
- The datasets build upon existing public datasets (FashionIQ and CIRR); users must also comply with the original licenses of those datasets.
- The model's performance relies heavily on the quality of instruction parsing, and real-world multi-modification accuracy may vary based on domain-specific data.

---

## 📝⭐️ Citation

If you find our paper, the M-FashionIQ/M-CIRR datasets, or this codebase useful in your research, please consider leaving a **Star** ⭐️ on our GitHub repo and citing our work:

```bibtex
@inproceedings{TEMA,
  title={TEMA: Anchor the Image, Follow the Text for Multi-Modification Composed Image Retrieval},
  author={Li, Zixu and Hu, Yupeng and Fu, Zhiheng and Chen, Zhiwei and Li, Yongqi and Nie, Liqiang},
  booktitle={Proceedings of the Association for Computational Linguistics (ACL)},
  year={2026}
}
```