ConceptEdit: Unlocking the Potential of Image Editing via Concept Scaling and Dense Supervision

   

ConceptEdit-12M is a large-scale image editing dataset. Each sample is stored as a triplet:

1. a source image,
2. an edited image,
3. a JSON metadata file describing the edit instruction, edit category, relative image paths, and VQA-style quality checks.

The dataset is packaged as multiple .tar shards. All paths inside the tar files and JSON files are relative paths; no local absolute paths are included.

Data splits

The released data contains four main splits:

File structure

After downloading, the dataset is organized as follows:

ConceptEdit-12M/
├── enhanced_prompt_square_resolution/
│   ├── batch_0.tar
│   ├── batch_1.tar
│   └── ...
├── enhanced_prompt_random_resolution/
│   ├── batch_0.tar
│   ├── batch_1.tar
│   └── ...
├── original_prompt_square_resolution/
│   ├── batch_0.tar
│   ├── batch_1.tar
│   └── ...
└── original_prompt_random_resolution/
    ├── batch_0.tar
    ├── batch_1.tar
    └── ...

Each tar shard contains one batch directory. For example:

batch_0/
├── 0_0_3.json
├── 0_0_3_source.jpg
├── 0_0_3_edit.png
├── 0_0_4.json
├── 0_0_4_source.jpg
├── 0_0_4_edit.png
└── ...

For each sample ID, the three files are:

<sample_id>.json
<sample_id>_source.jpg
<sample_id>_edit.png

The JSON file contains relative paths to the corresponding source and edited images, for example:

{
  "id": "0_0_3",
  "images": {
    "source": "batch_0/0_0_3_source.jpg",
    "edited": "batch_0/0_0_3_edit.png"
  }
}

How to extract

The following commands are provided for format reference only. You may use faster extraction tools or multi-threaded extraction methods according to your own needs.

Extract a single shard:

mkdir -p ConceptEdit-12M/extracted/enhanced_prompt_square_resolution

tar -xf ConceptEdit-12M/enhanced_prompt_square_resolution/batch_0.tar \
  -C ConceptEdit-12M/extracted/enhanced_prompt_square_resolution

List the content of a shard without extracting:

tar -tf ConceptEdit-12M/enhanced_prompt_square_resolution/batch_0.tar | head

Extract all shards in one split:

SPLIT=enhanced_prompt_square_resolution
mkdir -p ConceptEdit-12M/extracted/${SPLIT}

for f in ConceptEdit-12M/${SPLIT}/batch_*.tar; do
  tar -xf "$f" -C ConceptEdit-12M/extracted/${SPLIT}
done

Extract all four main splits:

for SPLIT in \
  enhanced_prompt_square_resolution \
  enhanced_prompt_random_resolution \
  original_prompt_square_resolution \
  original_prompt_random_resolution; do

  mkdir -p ConceptEdit-12M/extracted/${SPLIT}
  for f in ConceptEdit-12M/${SPLIT}/batch_*.tar; do
    tar -xf "$f" -C ConceptEdit-12M/extracted/${SPLIT}
  done

done

JSON format

Each sample has one JSON file with the following top-level fields:

id
images
edit_concept
instruction
evaluation
original_simple_caption

Field descriptions:

• id: sample ID. The image and JSON filenames are derived from this ID.
• images: relative paths to the source image and edited image.
  • source: source/original image path inside the extracted batch directory.
  • edited: edited image path inside the extracted batch directory.
• edit_concept: taxonomy information for the edit.
  • category
  • sub_category
  • task
  • detail
• instruction: natural-language edit instructions.
  • short_en
  • short_zh
  • detailed_en
  • detailed_zh
• evaluation: VQA-style filtering and quality-control metadata.
  • overall_vqa_score
  • keep
  • wrong_count
  • recaption_prompt_en
  • recaption_prompt_zh
  • vqa: a list of VQA checks, each containing:
    • dimension
    • question_en
    • question_zh
    • expected_answer
    • passed
• original_simple_caption: a short English caption for the source image.

A simplified example:

{
  "id": "0_0_3",
  "images": {
    "source": "batch_0/0_0_3_source.jpg",
    "edited": "batch_0/0_0_3_edit.png"
  },
  "edit_concept": {
    "category": "...",
    "sub_category": "...",
    "task": "...",
    "detail": "..."
  },
  "instruction": {
    "short_en": "...",
    "short_zh": "...",
    "detailed_en": "...",
    "detailed_zh": "..."
  },
  "evaluation": {
    "overall_vqa_score": 1.0,
    "keep": true,
    "wrong_count": 0,
    "recaption_prompt_en": "",
    "recaption_prompt_zh": "",
    "vqa": [
      {
        "dimension": "1_primary_edit_success",
        "question_en": "...",
        "question_zh": "...",
        "expected_answer": "yes",
        "passed": true
      }
    ]
  },
  "original_simple_caption": "A short English caption for the source image."
}

Source image acknowledgement

The source images in ConceptEdit-12M are based on images from Fine-T2I.

Citation

If you find this dataset useful, please cite:

@article{cui2026unlocking,
  title={Unlocking the Potential of Image Editing via Concept Scaling and Dense Supervision},
  author={Cui, Long and Liu, Xiaoqian and Qin, Qi and Xin, Yi and Lin, Tao and Li, Jianguo and Zhang, Linfeng},
  journal={arXiv preprint arXiv:2608.16812},
  year={2026}
}