---
license: apache-2.0
base_model: dima806/ai_vs_human_generated_image_detection
tags:
- image-classification
- vision
- ai-detection
- deepfake-detection
- vit
metrics:
- accuracy
- f1
pipeline_tag: image-classification
---

# CapCheck AI vs Human-Generated Image Detection

Vision Transformer (ViT) fine-tuned for detecting AI-generated images.
Uses newer training data than the original CIFAKE-based model.

## Model Lineage & Attribution

This model builds on the work of others:

| Layer | Model | Author | License |
|-------|-------|--------|---------|
| Base Architecture | [google/vit-base-patch16-224-in21k](https://huggingface.co/google/vit-base-patch16-224-in21k) | Google | Apache 2.0 |
| AI Detection Fine-tune | [dima806/ai_vs_human_generated_image_detection](https://huggingface.co/dima806/ai_vs_human_generated_image_detection) | dima806 | Apache 2.0 |
| This Model | capcheck/ai-human-generated-image-detection | CapCheck | Apache 2.0 |

**Special thanks to:**
- **Google** for the Vision Transformer (ViT) architecture
- **dima806** for fine-tuning on AI vs human-generated image datasets

## Model Description

- **Architecture**: ViT-Base (86M parameters)
- **Input Size**: 224x224 pixels
- **Training Data**: AI-generated vs human-created images (newer dataset than CIFAKE)
- **Task**: Binary classification (human vs AI-generated)

## Usage

```python
from transformers import pipeline

detector = pipeline("image-classification", model="capcheck/ai-human-generated-image-detection")
result = detector("path/to/image.jpg")

# Output:
# [{"label": "AI-generated", "score": 0.95}, {"label": "human", "score": 0.05}]
```

## Labels

| Label | Description |
|-------|-------------|
| `human` | Authentic photograph or human-created image |
| `AI-generated` | AI-generated or synthetically created image |

## Comparison with capcheck/ai-image-detection

| Feature | ai-image-detection | ai-human-generated-image-detection |
|---------|-------------------|-------------------------------------|
| Source | dima806/ai_vs_real_image_detection | dima806/ai_vs_human_generated_image_detection |
| Training Data | CIFAKE dataset | Newer AI vs human dataset |
| Labels | `REAL` / `FAKE` | `human` / `AI-generated` |
| Handle | `ai-image-detection` | `ai-human-generated-image-detection` |

## Limitations

- May have reduced accuracy on very new AI generators not in training data
- Heavily compressed images (low JPEG quality) can affect results
- Works best on images with clear subjects (224x224+ pixels)

## Intended Use

- Content moderation and authenticity verification
- Research into AI-generated content detection
- Educational purposes

**Not intended for**:
- Making consequential decisions without human review
- Law enforcement evidence without corroborating sources

## License

Apache 2.0 (inherited from Google ViT and dima806's fine-tuned model)

## Citation

```bibtex
@misc{capcheck-ai-human-generated-detection,
  author = {CapCheck},
  title = {AI vs Human-Generated Image Detection Model},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/capcheck/ai-human-generated-image-detection},
  note = {Based on dima806/ai_vs_human_generated_image_detection, fine-tuned from google/vit-base-patch16-224-in21k}
}
```