Dataset: The CoBSAT benchmark evaluates the ability of MLLMs to perform T2I-ICL. It covers five themes: color, background, style, action, and texture, each with two different emphases: object-inference and attribute-inference. Here, we visualize the images and their corresponding labels and captions collected for our dataset. We further integrate the images and their labels for constructing the prompts for text-to-image in-context learning using the processing code provided in https://github.com/UW-Madison-Lee-Lab/CoBSAT.

Paper Link: https://arxiv.org/abs/2402.01293

@article{zeng2024can,
  title={Can MLLMs Perform Text-to-Image In-Context Learning?},
  author={Zeng, Yuchen and Kang, Wonjun and Chen, Yicong and Koo, Hyung Il and Lee, Kangwook},
  journal={arXiv preprint arXiv:2402.01293},
  year={2024}
}