GPT-Image-Edit-1.5M A Million-Scale, GPT-Generated Image Dataset

📃Arxiv | 🌐 Project Page | 💻Github

GPT-Image-Edit-1.5M is a comprehensive image editing dataset that is built upon HQ-Edit, UltraEdit, OmniEdit and Complex-Edit, with all output images regenerated with GPT-Image-1.

📣 News

• [2025.08.20] 🚀 We provide a script for multi-process downloading. See Multi-process Download.
• [2025.07.27] 🤗 We release GPT-Image-Edit, a state-of-the-art image editing model with 1.5M high-quality editing samples. All data, models, training code and evaluation code are open-sourced. Our code is based on UniWorld-V1, Thanks to the authors of UniWorld-V1. Checking our report for more details. Welcome to watch 👀 this repository for the latest updates.

Dataset Statistics Summary

Full Dataset Overview

Instruction Complexity Distribution

• Original Instructions :
  • 1,140,182 samples
  • Same instructions from the original dataset
  • Basic to moderate complexity
• Rewrited Instructions:
  • 100,000 samples
  • Instructions rewrited based on the input image and new output image
  • Enhanced complexity
• Complex Instructions:
  • 313,393 samples
  • Complex-edit style instructions, with $C_3$ level complexity
  • Advanced complexity

Detailed Breakdown by Source and Task

HQ-Edit Dataset (183,182 samples)

Output images of 89,585 samples in edit subfolder are based on the original input images of HQ-Edit but edited with GPT-Image-1. 93,597 samples in generate subfolder have the input images re-genererated with GPT-Image-1 and then edited by the same model to produce output images.

OmniEdit Dataset (1,270,385 samples)

Output images from OmniEdit are re-edited with original input images and instructions with GPT-Image-1. Additionally, we sampled 313,393 input images from OmniEdit and generated Complex-Edit style instructions, with $C_3$ level complexity.

UltraEdit Dataset (100,008 samples)

100,008 samples from UltraEdit, uniformly sampled from 9 categories, have original input images re-edited by GPT-Image-1 with original instructions.

Unified Directory Structure

gpt-edit/
├── hqedit/
│   ├── edit/
│   │   ├── input/                    # Original input images
│   │   ├── output/                   # GPT-generated edited images
│   │   └── metadata/
│   │       └── hqedit_edit.json      # 89,585 samples
│   └── generate/
│       ├── input/                    # Generated input images
│       ├── output/                   # Generated output images
│       └── metadata/
│           └── hqedit_generate.json  # 93,597 samples
├── omniedit/
│   ├── addition/
│   │   ├── input/                    # Original input images
│   │   ├── output/                   # Original + GPT outputs
│   │   └── metadata/
│   │       └── omniedit_addition.json
│   ├── attribute_modification/
│   ├── background_swap/
│   ├── complex-edit/
│   ├── env/
│   ├── object_swap/
│   ├── removal/
│   ├── style/
│   └── swap/
└── ultraedit/
    ├── add/
    │   ├── input/                    # Original input images
    │   ├── output/                   # GPT-generated outputs
    │   └── metadata/
    │       └── ultraedit_add.json
    ├── change_color/
    ├── change_global/
    ├── change_local/
    ├── others/
    ├── replace/
    ├── transform_global/
    ├── transform_local/
    └── turn/

Please note that samples in gpt-edit/omniedit/swap are NOT a third kind of swap operation but haven't yet been classified into background_swap or object_swap.

Metadata Format

All metadata files follow a unified JSON structure:

Common Fields

{
  "id": "string",                    # Unique identifier: <dataset>_<task>_<id>
  "dataset_source": "string",       # "hqedit" | "omniedit" | "ultraedit"
  "task": "string",                 # Task category (e.g., "edit", "addition", "add")
  "input": "string",                # Relative path to input image: "input/<id>.png"
  "output": "string",               # Relative path to output image: "output/<id>.png"
  "instruction": "string",          # Editing instruction text
  "instruction_type": "string",     # "original" | "rewrite" | "complex"
  "input_description": "string"     # Description of the input image
}

Dataset-Specific Fields

OmniEdit

{
  "instruction_original": "string"  # Original instruction (for rewrite cases)
}

Usage Guide

Download

You can download the dataset using git lfs from the Hugging Face Hub:

git lfs install
git clone https://huggingface.co/datasets/UCSC-VLAA/GPT-Image-Edit-1.5M

Multi-process Download

We provide a script for faster download with multi-processing.

1. Download the script.

   wget https://huggingface.co/datasets/UCSC-VLAA/GPT-Image-Edit-1.5M/blob/main/download.sh
   

2. Use the script for multi-process downloading.

   bash download.sh -d <dataset_name> -o <your_directory>/gpt-edit -p <process_number>
   

   <dataset_name> should be one of hqedit/ultraedit/omniedit.

Prepare the Data

The annotation JSON files are located in UCSC-VLAA/gpt-image-edit-training/training_json.

To prepare a data.txt file for training (as mentioned in the associated GitHub repository), use the following format:

1. The first column is the root path to the image.
2. The second column is the corresponding annotation JSON file.
3. The third column indicates whether to enable the region-weighting strategy (we use false in our training setting).

An example data.txt for gpt-edit can be found in the GitHub repository, or an example structure is:

data/gpt-edit/hqedit/edit,training_json/hqedit_gpt_edit.json,false
data/gpt-edit/hqedit/generate,training_json/hqedit_gpt_generate.json,false
data/gpt-edit/omniedit,training_json/omniedit_gpt.json,false
data/gpt-edit/omniedit,training_json/omniedit_gpt_rewrite.json,false
data/gpt-edit/omniedit/complex-edit,training_json/complexedit_gpt.json,false
data/gpt-edit/ultraedit,training_json/ultraedit_gpt.json,false

Working with Image Paths

Paths in metadata are relative to the task directory

# Input: "input/00070858.png" -> hqedit/edit/input/00070858.png
# Output: "output/00070858.png" -> hqedit/edit/output/00070858.png

📊 Benchmarks

GEdit-EN-full

Complex-Edit

ImgEdit-Full

👍 Acknowledgement and Related Work

• UniWorld-V1: UniWorld-V1 is a unified framework for understanding, generation, and editing.
• ImgEdit: ImgEdit is a large-scale, high-quality image-editing dataset comprising 1.2 million carefully curated edit pairs and a comprehensive benchmark for image editing.
• Complex-edit: Complex-edit is benchmark for complex image editing.
• Qwen2.5-VL: The new flagship vision-language model of Qwen.
• FLUX.1-Kontext-dev: A state-of-the-art image editing model.
• Step1X-Edit: A state-of-the-art image editing model and a comprehensive benchmark for image editing.
• OmniGen2: A state-of-the-art image editing model and a comprehensive benchmark for image editing.

Citation

If you find our paper useful, please cite us with

@misc{wang2025gptimageedit15mmillionscalegptgeneratedimage,
      title={GPT-IMAGE-EDIT-1.5M: A Million-Scale, GPT-Generated Image Dataset}, 
      author={Yuhan Wang and Siwei Yang and Bingchen Zhao and Letian Zhang and Qing Liu and Yuyin Zhou and Cihang Xie},
      year={2025},
      eprint={2507.21033},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.21033}, 
}