AI vs. Real Image Classifier

A Google Colab notebook for training an image classifier that distinguishes AI-generated (FAKE) images from real (REAL) images. The project uses a pretrained EfficientNetV2-S model with PyTorch and saves training progress to Google Drive.

Live website and source code

Open the live AI vs. Real website →

Project	Repository	Purpose
Frontend	DNeberize/ai-or-real	Website interface for uploading images and viewing predictions
Backend	DNeberize/ai-vs-real-backend	API that loads the trained model and processes prediction requests
Model training	This repository	Google Colab workflow used to train and export the classifier

Complete project: this notebook trains the model, the backend serves it through an API, and the frontend provides the public user interface.

Important: this project runs in Google Colab

This notebook was written in Google Colab and is intended to run there. The current code depends on Colab-specific features:

• google.colab.drive.mount() for access to the dataset and saved checkpoints;
• google.colab.files.upload() for testing individual images;
• Colab paths such as /content and /content/drive/MyDrive;
• a Colab GPU runtime for practical training speed.

Because of these dependencies, the notebook will not run unchanged in a local Jupyter environment or directly on GitHub. GitHub stores and displays the notebook; Google Colab provides the runtime that executes it and authorizes access to Google Drive.

Google Drive can also be accessed outside Colab through the Google Drive API or other authentication methods, but that setup is not implemented in this project.

Project structure

ai_vs_real/
├── ai_vs_real.ipynb   # Training, evaluation, checkpointing, and inference
├── README.md             # Project documentation
└── .gitignore            # Excludes datasets, checkpoints, and local files

The dataset and trained model files are intentionally kept in Google Drive rather than in this GitHub repository.

Dataset setup

Create this folder in Google Drive:

My Drive/ai_vs_real_images/

Place the dataset archive here:

My Drive/ai_vs_real_images/Compressed.zip

The ZIP archive should contain this structure. TRAIN is required; VAL and TEST are optional:

Compressed.zip
├── TRAIN/
│   ├── FAKE/
│   └── REAL/
├── VAL/                  # Optional; otherwise 20% of TRAIN is used
│   ├── FAKE/
│   └── REAL/
└── TEST/                 # Optional, but recommended for final evaluation
    ├── FAKE/
    └── REAL/

An extra top-level folder inside the ZIP is supported because the notebook automatically flattens one nested directory after extraction.

How to run

1. Upload Compressed.zip to the Google Drive path shown above.
2. Open ai_vs_real.ipynb in Google Colab.
   • In Colab, select File → Open notebook → GitHub and enter this repository's URL.
   • You can also upload the notebook directly to Colab.
3. Select Runtime → Change runtime type → GPU.
4. Run the notebook cells from top to bottom.
5. Approve the Google Drive permission request when prompted.

The first model run may download pretrained EfficientNet weights. Standard Colab environments normally include the required packages: PyTorch, TorchVision, NumPy, Pillow, scikit-learn, and tqdm.

What the notebook does

• mounts Google Drive and extracts the dataset into the temporary Colab runtime;
• caches dataset information to avoid unnecessary extraction;
• creates a stratified 80/20 training-validation split when a VAL folder is not supplied;
• applies image augmentation and ImageNet normalization;
• fine-tunes an EfficientNetV2-S binary classifier;
• uses mixed-precision training when a GPU is available;
• tracks training and validation loss, accuracy, and macro F1 score;
• saves resumable checkpoints and stops early when validation F1 stops improving;
• evaluates the best checkpoint on TEST, when present;
• predicts an uploaded image as AI-Generated or Real.

Main configuration

The main settings are near the beginning of the notebook:

Setting	Default	Purpose
RESUME_FROM_LAST	True	Continue the latest saved run
TARGET_TOTAL_EPOCHS	12	Total number of training epochs
VAL_FRACTION	0.20	Validation fraction when VAL is absent
RANDOM_SEED	42	Reproducible data split and training seed
ENABLE_JPEG_AUG	True	Apply low-probability JPEG compression augmentation
TRAIN_IMAGE_SIZE	224	Input image size
TRAIN_BATCH_SIZE	128	Training batch size
EVAL_BATCH_SIZE	256	Validation and test batch size
BACKBONE_NAME	efficientnet_v2_s	Pretrained model architecture

If Colab runs out of GPU memory, reduce TRAIN_BATCH_SIZE and EVAL_BATCH_SIZE.

Saved files

Training artifacts are written to:

My Drive/ai_vs_real_model_checkpoints/

The notebook creates files similar to:

ai_vs_real_model_checkpoints/
├── best_model.pth
├── dataset_cache.json
├── latest_run.json
└── runs/<timestamp>/
    ├── best_checkpoint.pth
    ├── last_checkpoint.pth
    ├── epoch_metrics.csv
    ├── epoch_metrics.jsonl
    ├── run_config.json
    └── run_summary.json

With RESUME_FROM_LAST = True, rerunning the notebook continues from last_checkpoint.pth. Set it to False when you want to create a new training run.

Inference

After training, run the inference cells at the end of the notebook. Colab will open a file picker, load the best available checkpoint from Google Drive, and print the predicted label and confidence score.

Model predictions are experimental and should not be treated as definitive proof that an image is real or AI-generated.