tianleliphoebe/DreamEditBench

DreamEditBench for Subject Replacement task and Subject Addition task.

Dataset Description

• —Homepage: https://dreameditbenchteam.github.io
• —Repository: https://github.com/DreamEditBenchTeam/DreamEdit <!-- Paper: https://arxiv.org/abs/2306.12624 -->

The goal of subject replacement is to replace a subject from a source image with a customized subject. In contrast, the aim of the subject addition task is to add a customized subject to a desired position in the source image. To standardize the evaluation of the two proposed tasks, we curate a new benchmark, i.e. DreamEditBench, consisting of 22 subjects in alignment with DreamBooth with 20 images for each subject correspondingly. For the subject replacement task, we collect 10 images for each type, which include same-typed source subjects in diverse environments. The images are retrieved from the internet with the search query “a photo of [Class name]”, and the source subject should be the main subject in the image which dominates a major part of the photo. For the subject addition task, we collect 10 reasonable backgrounds for each type of subject. In the meantime, we manually designate the specific location the target subject should be placed with a bounding box in the background. To collect the specific backgrounds for each subject, we first brainstorm and list the possible common environments of the subjects, then we search the listed keywords from the internet to retrieve and pick the backgrounds

Data Structure

There are 22 subject folders in each task folder respectively. In each subject folder, there are 10 source images. For Subject Addition task, there is an additional bbox.json file recording the manually labeled bounding box for each background. The replacementsubset.csv and additionsubset.csv record the easy/hard subset division for each task correspondingly.

Citation Information

If you find this dataset useful, please consider citing our paper:

@misc{li2023dreamedit,
      title={DreamEdit: Subject-driven Image Editing}, 
      author={Tianle Li and Max Ku and Cong Wei and Wenhu Chen},
      year={2023},
      eprint={2306.12624},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}