🥭 NanoMango: AI-Generated Image Detector

NanoMango is a Deep Learning project designed to distinguish between Real photographs and Fake (AI-generated) images specifically from Google Gemini 2.5 Flash Image Preview. It utilizes Transfer Learning with a MobileNetV3 Large architecture to achieve high accuracy with a lightweight footprint.

The project includes a training engine and a user-friendly web interface (Streamlit) for live testing but you have to run it locally.

⚠️ Note: This detector is trained specifically on images generated by NanoBanana (Google Gemini 2.5 Flash Image Preview). Performance on other AI-generated images or models may vary.

Preview

Inference:

Training:

Project Structure

NanoMango/
├── data/                  # Dataset folder (not included in repo)
│   ├── real/              # Real images from Open Images V7 dataset
│   └── fake/              # AI-generated images(not the pro for now)
├── app.py                 # Streamlit Web Interface for testing
├── train.py               # Main training script
├── environment.yml        # Conda environment dependencies
├── nanomango_best.pth     # Saved model weights (generated after training)
└── README.md              # Project documentation

🚀 Installation & Setup

This project uses a generic Conda environment to manage dependencies (PyTorch, Streamlit, Torchvision, etc.).

1. Clone the repository:

git clone <your-repo-url>
cd NanoMango

1. Create the Environment: Ensure you have Conda installed, then run:

conda env create -f environment.yml

1. Activate the Environment:

conda activate nanomango

🧠 Training the Model

To train the model from scratch, ensure your data is organized in the data/ folder with two subdirectories: real and fake.

1. Run the training script:

python train.py

1. Process:

• The script automatically detects your hardware (CUDA, Apple Silicon MPS, or CPU).
• It fine-tunes a pre-trained MobileNetV3 Large model.
• The best weights are saved automatically as nanomango_best.pth.

Model Configuration:

• Base: MobileNetV3 Large (Pre-trained on ImageNet)
• Input Size: 224x224 (Resized)
• Classes: * 0: Fake
• 1: Real (Note: PyTorch assigns these alphabetically: F comes before R)

🧪 Running the Web Demo

You can test the trained model using the built-in web interface.

1. Start the App:

streamlit run app.py

1. Use the Interface:

• A browser tab will open at http://localhost:8501.
• Step 1: Upload your nanomango_best.pth file.
• Step 2: Upload any image (JPEG/PNG) to test.
• The app will display the prediction and confidence score in real-time.

🛠 Tech Stack

• Core Framework: PyTorch
• Architecture: MobileNetV3 Large
• Interface: Streamlit
• Image Processing: Torchvision & Pillow

Acknowledgments

This project uses datasets from the following sources:

• Real Images: Open Images V7 - A large-scale dataset of real photographs curated by Bitmind
• Fake Images: NanoBanana - AI-generated images from the NanoBanana model, provided by Bitmind

Thank you to the open-source community and Bitmind for making these datasets publicly available on Hugging Face.

License

Apache License