README.md

metadata

license: mit
task_categories:
  - text-to-image
  - image-to-text
language:
  - en
tags:
  - art
  - anime
  - safebooru
  - stable-diffusion
  - flux
pretty_name: Safebooru Anime Dataset
size_categories:
  - n<1K
dataset_info:
  features:
    - name: image
      dtype: image
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 328858750
      num_examples: 389
  download_size: 324178997
  dataset_size: 328858750
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Safebooru Anime Dataset

This dataset contains 389 curated anime-style images from Safebooru with descriptive captions, prepared for fine-tuning text-to-image models like Stable Diffusion and FLUX.

Dataset Structure

Each example in the dataset contains:

• image: The anime artwork image
• text: Descriptive caption/tags for the image
• file_name: The filename of the image
• original_post_id: Original Safebooru post ID
• original_filename: Original filename from Safebooru

Usage

Loading the Dataset

from datasets import load_dataset

# Load from HuggingFace Hub
dataset = load_dataset("your-username/your-dataset-name")

# Access images and captions
for example in dataset['train']:
    image = example['image']  # PIL Image
    caption = example['text']  # String caption
    print(f"Caption: {caption}")
    image.show()
    break

For Fine-tuning

This dataset is formatted for fine-tuning text-to-image models:

from datasets import load_dataset
from transformers import pipeline

dataset = load_dataset("your-username/your-dataset-name")

# Use with your training script
train_dataset = dataset['train']

For FLUX LoRA Training

Compatible with the FLUX QLora training notebook: https://huggingface.co/blog/flux-qlora

Dataset Statistics

• Total images: 389
• Image formats: JPG, PNG
• Caption style: Booru-style tags in comma-separated format
• Average caption length: ~50-100 tokens

Data Preprocessing

Images were:

1. Downloaded from Safebooru with safe content filters
2. Resized to maintain quality while being training-friendly
3. Paired with descriptive captions converted from booru tags

License

MIT License - See individual image metadata for specific attribution requirements.

Citation

If you use this dataset, please cite:

@dataset{safebooru_anime_dataset,
  title={Safebooru Anime Dataset},
  author={Your Name},
  year={2024},
  publisher={HuggingFace}
}