Subjects200K is a large-scale dataset containing 200,000 paired images, introduced as part of the OminiControl project. Each image pair maintains subject consistency while presenting variations in scene context.

Quick Start

• Load dataset

  from datasets import load_dataset
  
  # Load dataset
  dataset = load_dataset('Yuanshi/Subjects200K')
  

• Filter high-quality pairs from collection_2

  def filter_func(item):
      if item.get("collection") != "collection_2":
          return False
      if not item.get("quality_assessment"):
          return False
      return all(
          item["quality_assessment"].get(key, 0) >= 5
          for key in ["compositeStructure", "objectConsistency", "imageQuality"]
      )
  
  collection_2_valid = dataset["train"].filter(
      filter_func,
      num_proc=16,
      cache_file_name="./cache/dataset/collection_2_valid.arrow", # Optional
  )
  

Collections

Collection1 (collection_1)

• 512 x 512 resolution, with 16-pixel padding.
• Total 18,396 image pairs, with 8,200 pairs having high-quality ratings.

Collection2 (collection_2)

• 512 x 512 resolution, with 16-pixel padding.
• Total 187,840 image pairs, with 111,767 pairs having high-quality ratings.

Collection3 (collection_3) link

• 1024 x 1024 resolution.

The description formats may vary across different collections.

Data Format

Contributing

We welcome contributions! Please feel free to submit a Pull Request or open an Issue.

Citation

@article{
  tan2024omini,
  title={OminiControl: Minimal and Universal Control for Diffusion Transformer},
  author={Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang},
  journal={arXiv preprint arXiv:2411.15098},
  year={2024}
}

This repo contains the dataset used in OminiControl: Minimal and Universal Control for Diffusion Transformer.