InstanceControl Datasets (MIG-Train / MIG-Eval)

This dataset is part of the work presented in the paper InstanceControl: Controllable Complex Image Generation without Instance Labeling (ECCV 2026).

• Paper: arXiv:2606.31924
• Project Page: InstanceControl Homepage
• Repository: GitHub - InstanceControl

Dataset Description

MIG-Train and MIG-Eval are designed for multi-instance controllable generation without requiring manual instance labeling.

• MIG-Train contains training annotations built from images, captions, instance masks, and visual condition maps. It utilizes images from the Segment Anything Dataset, COCO 2017, and UniWorld-V1 / BLIP3o-60k.
• MIG-Eval is the benchmark dataset used to evaluate model performance on tasks requiring fine-grained control, measuring metrics such as IoU, local CLIP, and ACC.

For detailed instructions on data preparation, training, and evaluation, please refer to the GitHub repository.

Citation

@article{instancecontrol,
  title   = {InstanceControl: Controllable Complex Image Generation without Instance Labeling},
  author  = {Xiaoyu Liu and Huan Wang and Fan Li and Zhixin Wang and Jiaqi Xu and Ming Liu and Wangmeng Zuo},
  journal = {arXiv preprint arXiv:2606.31924},
  year    = {2026}
}