README.md

---
size_categories:
  - 1K<n<10K
license: other
tags:
- ghibli
- ai-generated
- image-classification
language:
- en
pretty_name: Ghibli Real vs AI Dataset
task_categories:
- image-classification
task_ids:
- multi-class-classification
splits:
  - name: train
    num_examples: 4257
annotations_creators:
- machine-generated
source_datasets:
  - Nechintosh/ghibli
  - nitrosocke/Ghibli-Diffusion
  - KappaNeuro/studio-ghibli-style
dataset_info:
  labels:
  - real
  - ai
---

# Ghibli Real vs AI-Generated Dataset

- One sample per line
- Includes: `id`, `image`, `label`, `description`
- Use this for standard classification or image-text training

- Real images sourced from [Nechintosh/ghibli](https://huggingface.co/datasets/Nechintosh/ghibli) (810 images)
- AI-generated images created using:
  - [nitrosocke/Ghibli-Diffusion](https://huggingface.co/nitrosocke/Ghibli-Diffusion) (2637 images)
  - [KappaNeuro/studio-ghibli-style](https://huggingface.co/KappaNeuro/studio-ghibli-style) (810 images)
  - Note: While the KappaNeuro repository does not explicitly state a license, it is a fine-tuned model based on [Stable Diffusion XL](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0), which is distributed under the CreativeML Open RAIL++-M License. Therefore, it is assumed that this model inherits the same license and non-commercial restrictions.

### How to load

```python
from datasets import load_dataset

samples = load_dataset("pulnip/ghibli-dataset", split="train")

# Convert labels to binary classification: 'real' vs 'ai'
# Note: The original "label" field contains "real", "nitrosocke", and "KappaNeuro".
#       You can treat all non-"real" labels as "ai" to use this dataset for binary classification.
for sample in samples:
    sample["binary_label"] = "real" if sample["label"] == "real" else "ai"
```

## License and Usage

This dataset combines data from multiple sources. Please review the licensing conditions carefully.

### Real Images
- Source: [Nechintosh/ghibli](https://huggingface.co/datasets/Nechintosh/ghibli)
- License: Not explicitly stated; assumed for **non-commercial research use only**

### AI-Generated Images
- Source models:
  - [nitrosocke/Ghibli-Diffusion](https://huggingface.co/nitrosocke/Ghibli-Diffusion)  
    Loaded with: `torch_dtype=torch.float16`
  - [KappaNeuro/studio-ghibli-style](https://huggingface.co/KappaNeuro/studio-ghibli-style)  
    Loaded with: `torch_dtype=torch.float16, variant="fp16"`
- These models are provided under community licenses that generally restrict usage to **non-commercial and research purposes**.

---

### Summary

This repository is not published under a single license such as MIT.  
Because the dataset includes content from multiple sources with varying restrictions,  
**the dataset is licensed under 'other' and should be treated as non-commercial research-use only.**

Users are responsible for reviewing each component’s license terms before redistribution or adaptation.