README.md

metadata

license: cc-by-nc-4.0
task_categories:
  - image-to-image
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Note: FineEdit preview subset (1k samples) is now available. Full dataset will be released within July.

Bounding Boxes

This dataset provides two bounding box annotations for each sample:

• bbox: Obtained from an object detection tool, indicating the target region in the source image.
• bbox_edit: Derived by computing the pixel difference between the source and edited images, followed by refinement with traditional computer vision algorithms.

Citation

@misc{xu2026fineeditfinegrainedimageedit,
  title={FineEdit: Fine-Grained Image Edit with Bounding Box Guidance},
  author={Haohang Xu and Lin Liu and Zhibo Zhang and Rong Cong and Xiaopeng Zhang and Qi Tian},
  year={2026},
  eprint={2604.10954},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2604.10954},
}