inclusionAI/ConceptEdit-Bench

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

     

ConceptEdit-Bench is the evaluation benchmark for ConceptEdit. It contains 1,000 curated image editing test cases across 6 major editing categories. Each test case includes a source image and a JSON metadata file with the edit instruction, taxonomy information, and a relative path to the source image.

The benchmark evaluation code is provided in the ConceptEdit GitHub repository:

https://github.com/inclusionAI/ConceptEdit

Files

This Hugging Face dataset repository provides the benchmark data package:

conceptbench_data.tar

The tar file contains one top-level directory:

data/

How to extract

Extract the benchmark package with:

mkdir -p ConceptEdit-Bench

tar -xf conceptbench_data.tar -C ConceptEdit-Bench

After extraction, the expected layout is:

ConceptEdit-Bench/
└── data/
    ├── taxonomy.json
    ├── images/
    │   ├── <image_id>.jpg
    │   └── ...
    ├── advanced_domain_application/
    ├── general_object_editing/
    ├── generation_composition/
    ├── global_enhancement_atmosphere/
    ├── portrait_human_specialized/
    └── text_graphic_design/

Each benchmark case JSON is stored under the taxonomy hierarchy:

data/<category>/<sub_category>/<task>/<detail>.json

The source images are stored under:

data/images/

You can inspect the tar file without extracting it:

tar -tf conceptbench_data.tar | head

Expected data size

Item	Count
Case JSON files	1,000
Source images	979
Top-level categories	6

Some source images are shared by multiple benchmark cases, so the number of source images is smaller than the number of case JSON files.

JSON format

Each case JSON keeps only the fields needed for benchmark inference and evaluation:

{
  "caption": "source image caption",
  "edit_concept": {
    "category": "...",
    "sub_category": "...",
    "task": "...",
    "detail": "..."
  },
  "instruction_en": "short English edit instruction",
  "instruction_zh": "short Chinese edit instruction",
  "detailed_instruction_en": "detailed English edit instruction",
  "detailed_instruction_zh": "detailed Chinese edit instruction",
  "local_image_path": "images/example.jpg"
}

Field descriptions:

• caption: short English caption for the source image.
• edit_concept: taxonomy information for the editing case.
• instruction_en / instruction_zh: short English and Chinese edit instructions.
• detailed_instruction_en / detailed_instruction_zh: detailed English and Chinese edit instructions.
• local_image_path: relative path to the source image under the extracted data/ directory.

For example, if a JSON file contains:

{
  "local_image_path": "images/example.jpg"
}

then the corresponding source image should be located at:

ConceptEdit-Bench/data/images/example.jpg

All paths inside the released JSON files are relative paths. No local absolute paths are included.

Using with the evaluation code

Clone the ConceptEdit code repository and place the extracted data/ directory under the benchmark code directory expected by the scripts:

git clone https://github.com/inclusionAI/ConceptEdit.git
cd ConceptEdit/benchmark

# place or symlink the extracted data directory here
# expected path: ConceptEdit/benchmark/data/

The benchmark scripts read cases from ./data and use local_image_path to find source images.

Source Image Acknowledgement

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

Citation

If you find this benchmark useful, please cite the ConceptEdit paper and refer to the project repository:

@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}
}