CSIP (Contrastive Anime Style Image Pre-Training) Dataset

Summary

The CSIP (Contrastive anime Style Image Pre-Training) dataset is a comprehensive collection of anime-style images designed for style classification and artist recognition tasks. This dataset contains a diverse range of artwork from various anime artists, making it particularly valuable for computer vision applications focused on artistic style analysis and content-based image retrieval. The dataset is organized to facilitate contrastive learning approaches where models can learn to distinguish between different artistic styles and identify characteristic features of individual artists.

This dataset represents the uncleaned raw version of the CSIP collection, providing researchers and developers with access to the original, unprocessed source material. The images are distributed across multiple zip archives (dataset_p0.zip through dataset_p12.zip) to facilitate easier downloading and management. The dataset's primary focus on anime art styles makes it particularly useful for training models that need to understand and classify different artistic approaches within the anime genre, from character design to background art and composition techniques.

With a size category of 1M<n<10M, this dataset offers substantial scale for training robust deep learning models while maintaining manageable storage requirements through its partitioned structure. The dataset supports zero-shot classification tasks, enabling models to recognize and classify styles from artists not seen during training. This makes it valuable for research in transfer learning and few-shot learning scenarios where model generalization across different artistic styles is crucial.

The CSIP dataset exists in multiple versions to accommodate different use cases and quality requirements. Researchers can choose between the raw uncleaned version (this repository), a roughly cleaned version, or a human-curated evaluation set depending on their specific needs for training data quality and annotation reliability.

Dataset Structure

The dataset is organized into 13 zip archive files:

• dataset_p0.zip through dataset_p12.zip

Each archive contains a portion of the overall dataset, allowing for flexible downloading and storage management.

Usage

This dataset is provided in its raw, uncleaned format. Users should be prepared to perform their own data preprocessing, cleaning, and validation based on their specific use cases.

Related Datasets

• Uncleaned Raw Version: deepghs/csip (this repository)
• Roughly Cleaned Version: deepghs/csip_v1
• Human Picked Evaluation Set: deepghs/csip_eval (smaller, suitable for evaluation purposes)

Citation

@misc{csip_dataset,
  title        = {CSIP (Contrastive Anime Style Image Pre-Training) Dataset},
  author       = {deepghs},
  howpublished = {\url{https://huggingface.co/datasets/deepghs/csip}},
  year         = {2023},
  note         = {Uncleaned raw version containing anime artwork from various artists for style classification and artist recognition tasks},
  abstract     = {The CSIP (Contrastive anime Style Image Pre-Training) dataset is a comprehensive collection of anime-style images designed for style classification and artist recognition tasks. This dataset contains a diverse range of artwork from various anime artists, making it particularly valuable for computer vision applications focused on artistic style analysis and content-based image retrieval. The dataset is organized to facilitate contrastive learning approaches where models can learn to distinguish between different artistic styles and identify characteristic features of individual artists. This version represents the uncleaned raw collection distributed across multiple zip archives.},
  keywords     = {anime, style-classification, art, computer-vision, contrastive-learning}
}