DiffTell: A High-Quality Dataset for Describing Image Manipulation Changes (ICCV2025)

This repository contains the dataset and dataloader in the paper "DiffTell: A High-Quality Dataset for Describing Image Manipulation Changes".

Zonglin Di¹², Jing Shi¹, Yifei Fan¹, Hao Tan¹, Alexander Black³, John Collomosse¹³, Yang Liu^1
1Adobe Research ²University of California, Santa Cruz ³CVSSP, University of Surrey

📜 Paper | 📚 Dataset

Usage

The benchmark file structure should be like:

all_images/
  ├── 003157655.jpg
  ├── 003157655_res2.jpg
  ├── 003158361.jpg 
  ├── 003158361_res1.jpg
  ├── 004276100.jpg
  ├── 004276100_res2.jpg
...

One example of the dataset annotation is

[
    {
        "id": "42e7a57e-ef23-490c-bae8-c837d6ff8c07-swap",
        "images": [
            "all_images/005686983_res1.jpg",
            "all_images/005686983.jpg"
        ],
        "conversation": [
            {
                "role": "user",
                "content": "What are the differences between the two images?"
            },
            {
                "role": "assistant",
                "content": "The difference is in that add the text brizzly."
            }
        ],
        "source": "mario-10m"
    }
]

The way to load the dataset is given in diffTell_loader.py

Citation

If you find this helpful in your work, please consider citing our paper:

@inproceedings{di2025difftell,
  title={DiffTell: A High-Quality Dataset for Describing Image Manipulation Changes},
  author={Di, Zonglin and Shi, Jing and Fan, Yifei and Tan, Hao and Black, Alexander and Collomosse, John and Liu, Yang},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={24580--24590},
  year={2025}
}

Contact

If you have any questions regarding this work, please send email to [email protected].