---
task_categories: [image-classification]
language: [en]
license: cc-by-4.0
size_categories: [100K<n<1M]
tags: [image-classification, binary-classification, computer-vision, anime, 3d]

---

# Anime Blurry Preparation Dataset

## Summary

The **Anime Blurry Preparation Dataset** is a specialized **binary classification** dataset designed for distinguishing between clear and blurry 3D anime-style images. This dataset provides a comprehensive collection of image pairs where each clear image has a corresponding blurry counterpart, making it ideal for training and evaluating image quality assessment models. The dataset employs **automatic annotation** techniques using BLIP (Bootstrapping Language-Image Pre-training) to generate reliable labels for the classification task.

This dataset specifically focuses on **3D anime-style** content, addressing the unique challenges of image quality assessment in computer-generated anime imagery. The clear images represent high-quality, sharp renderings while the blurry counterparts simulate various types of image degradation that commonly occur during rendering, compression, or transmission processes. The **binary classification** framework makes this dataset particularly useful for developing models that can automatically detect and filter out low-quality images in anime content pipelines.

The dataset organization follows a structured approach with separate directories for clear and blurry images, each containing thousands of carefully curated samples. With a total size of approximately 5GB, this dataset provides substantial training data for developing robust image quality assessment models. The **automatic annotation** process ensures consistent labeling while maintaining the dataset's scalability and reproducibility for research and development purposes.

## Dataset Structure

The dataset is organized in a tar archive with the following structure:

- `clear/` - Directory containing clear, high-quality 3D anime images
- `blurry/` - Directory containing blurry versions of the same images

Each image file follows a consistent naming convention with hash-based identifiers to ensure uniqueness and prevent duplicates.

## Dataset Statistics

- **Total Size**: 5,088,880,640 bytes (approximately 5GB)
- **File Format**: JPEG images
- **Organization**: Binary classification (clear vs blurry)
- **Content Type**: 3D anime-style images
- **Annotation Method**: Automatic annotation using BLIP

## Original Content

binary classification dataset (clear-blurry) for 3d images,
auto-annotated by blip, just for testing, not recommended for production model training

## Citation

```bibtex
@misc{anime_blurry_preparation,
  title        = {Anime Blurry Preparation Dataset},
  author       = {deepghs},
  howpublished = {\url{https://huggingface.co/datasets/deepghs/anime_blurry_preparation}},
  year         = {2023},
  note         = {Binary classification dataset for clear vs blurry 3D anime images with automatic BLIP annotation},
  abstract     = {The Anime Blurry Preparation Dataset is a specialized binary classification dataset designed for distinguishing between clear and blurry 3D anime-style images. This dataset provides a comprehensive collection of image pairs where each clear image has a corresponding blurry counterpart, making it ideal for training and evaluating image quality assessment models. The dataset employs automatic annotation techniques using BLIP to generate reliable labels for the classification task. This dataset specifically focuses on 3D anime-style content, addressing the unique challenges of image quality assessment in computer-generated anime imagery. The clear images represent high-quality, sharp renderings while the blurry counterparts simulate various types of image degradation that commonly occur during rendering, compression, or transmission processes.},
  keywords     = {binary classification, image-classification, computer-vision, anime, 3d}
}
```