Subjects200K Dataset

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

Collection2 (collection_2)

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

Collection3 (collection_3) (To be uploaded)

• 1024 $\times$ 1024 resolution.

The description formats may vary across different collections.

Data Format

Key name	Type	Description
image	image	A composite image containing a pair of images with 16-pixel padding.
collection	str	The name or identifier of the source collection.
quality_assessment	dict	Quality scores evaluated by the ChatGPT-4o model. Each quality dimension is rated on a scale of 0-5. Recommended for filtering out low-quality image pairs which do not have all dimensions rated above 5.
description	dict	Detailed textual description of the image pair contents.

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}
}