|
|
--- |
|
|
license: apache-2.0 |
|
|
base_model: dima806/ai_vs_real_image_detection |
|
|
tags: |
|
|
- image-classification |
|
|
- vision |
|
|
- ai-detection |
|
|
- deepfake-detection |
|
|
- vit |
|
|
datasets: |
|
|
- CIFAKE |
|
|
metrics: |
|
|
- accuracy |
|
|
- f1 |
|
|
pipeline_tag: image-classification |
|
|
--- |
|
|
|
|
|
# CapCheck AI Image Detection |
|
|
|
|
|
Vision Transformer (ViT) fine-tuned for detecting AI-generated images. |
|
|
|
|
|
## 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_real_image_detection](https://huggingface.co/dima806/ai_vs_real_image_detection) | dima806 | Apache 2.0 | |
|
|
| This Model | capcheck/ai-image-detection | CapCheck | Apache 2.0 | |
|
|
|
|
|
**Special thanks to:** |
|
|
- **Google** for the Vision Transformer (ViT) architecture |
|
|
- **dima806** for fine-tuning on the CIFAKE dataset for AI image detection |
|
|
|
|
|
## Model Description |
|
|
|
|
|
- **Architecture**: ViT-Base (86M parameters) |
|
|
- **Input Size**: 224x224 pixels |
|
|
- **Training Data**: CIFAKE dataset (AI-generated vs real images) |
|
|
- **Task**: Binary classification (Real vs Fake/AI-generated) |
|
|
|
|
|
## Usage |
|
|
|
|
|
```python |
|
|
from transformers import pipeline |
|
|
|
|
|
detector = pipeline("image-classification", model="capcheck/ai-image-detection") |
|
|
result = detector("path/to/image.jpg") |
|
|
|
|
|
# Output: |
|
|
# [{"label": "Fake", "score": 0.95}, {"label": "Real", "score": 0.05}] |
|
|
``` |
|
|
|
|
|
## Labels |
|
|
|
|
|
| Label | Description | |
|
|
|-------|-------------| |
|
|
| `Real` | Authentic photograph or real-world image | |
|
|
| `Fake` | AI-generated or synthetically created image | |
|
|
|
|
|
## Performance |
|
|
|
|
|
This model was trained on the CIFAKE dataset. Performance on modern AI generators |
|
|
(Flux, Midjourney v6, DALL-E 3, Stable Diffusion 3) may vary. |
|
|
|
|
|
See [dima806's model card](https://huggingface.co/dima806/ai_vs_real_image_detection) |
|
|
for detailed training metrics. |
|
|
|
|
|
## Limitations |
|
|
|
|
|
- Trained primarily on older AI generators (pre-2024) |
|
|
- May have reduced accuracy on: |
|
|
- Very new AI generators not in training data |
|
|
- Heavily compressed images (low JPEG quality) |
|
|
- Images smaller than 224x224 pixels |
|
|
- Works best on images with clear subjects |
|
|
|
|
|
## 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 |
|
|
|
|
|
## Ethical Considerations |
|
|
|
|
|
- This tool is not 100% accurate - false positives harm legitimate creators |
|
|
- False negatives can allow misinformation to spread |
|
|
- Use in conjunction with other verification methods |
|
|
- Human review is recommended for high-stakes decisions |
|
|
|
|
|
## Roadmap |
|
|
|
|
|
### Current Version (v1.0.0) |
|
|
|
|
|
Base model from dima806's CIFAKE-trained ViT. Solid foundation for AI detection. |
|
|
|
|
|
### Planned Improvements |
|
|
|
|
|
**Phase 1: Modern Generator Training** |
|
|
- Fine-tune on images from Flux, Midjourney v6, DALL-E 3, Stable Diffusion 3 |
|
|
- Target: Reduce false negatives on 2024+ AI generators |
|
|
|
|
|
**Phase 2: False Positive Reduction** |
|
|
- Curate dataset of real images commonly flagged as AI |
|
|
- Photography edge cases: HDR, heavy editing, digital art |
|
|
- Target: <5% false positive rate |
|
|
|
|
|
**Phase 3: Continuous Updates** |
|
|
- Quarterly re-training as new generators emerge |
|
|
- Community feedback integration |
|
|
- Benchmark against latest AI generators |
|
|
|
|
|
### Contributing |
|
|
|
|
|
We welcome: |
|
|
- Dataset contributions (properly licensed images) |
|
|
- Bug reports and false positive/negative examples |
|
|
- Benchmark results on new generators |
|
|
|
|
|
Join the discussion: https://huggingface.co/capcheck/ai-image-detection/discussions |
|
|
|
|
|
## License |
|
|
|
|
|
Apache 2.0 (inherited from Google ViT and dima806's fine-tuned model) |
|
|
|
|
|
## Citation |
|
|
|
|
|
If you use this model, please cite: |
|
|
|
|
|
```bibtex |
|
|
@misc{capcheck-ai-detection, |
|
|
author = {CapCheck}, |
|
|
title = {AI Image Detection Model}, |
|
|
year = {2024}, |
|
|
publisher = {HuggingFace}, |
|
|
url = {https://huggingface.co/capcheck/ai-image-detection}, |
|
|
note = {Based on dima806/ai_vs_real_image_detection, fine-tuned from google/vit-base-patch16-224-in21k} |
|
|
} |
|
|
``` |
|
|
|
|
|
## Changelog |
|
|
|
|
|
### v1.0.0 (Initial Release) |
|
|
|
|
|
- Published base model from dima806/ai_vs_real_image_detection |
|
|
- Added proper attribution and documentation |
|
|
- Established as CapCheck's source of truth for AI image detection |
|
|
|