---
title: Detecting AI Generated Images & Model Attribution
emoji: 🖼️
colorFrom: indigo
colorTo: blue
sdk: streamlit
sdk_version: 1.39.0
app_file: streamlit_app.py
pinned: false
---

# Two-Stage AI Image Detection (HF Space)

This Space hosts a Streamlit app that performs:
1. Binary classification: real vs fake.
2. Origin classification (if fake): platform/source.

## Models
Instead of bundling large weight files (Space storage limit), the app downloads models at runtime from Hugging Face Hub using repo IDs:

```
Binary (real vs fake):    KomaAl/level_1_binary_model
Multiclass (origin):      KomaAl/level_2_multiclass_model
```

If you fork this Space, replace these IDs in `streamlit_app.py` with your own.

## Running Locally
```bash
streamlit run streamlit_app.py
```
First run may take longer while models download.

## Override Rule (If Implemented Externally)
If you later add a custom override for a particular label logic (e.g., Imagen confidence rule), integrate it inside `CascadeClassifier.predict` after computing multiclass probs.

## Requirements
See `requirements.txt` for pinned versions suitable for CPU inference.

## Notes
- Mixed precision (AMP) auto-enables on CUDA if available.
- To force CPU, modify `DEFAULT_DEVICE` in `streamlit_app.py`.
- Large model weights aren't stored in-repo; they're pulled via repo IDs to stay under Space storage limits.