Synthbuster

Implementation of Synthbuster: Towards Detection of Diffusion Model Generated Images.

Synthbuster is a tool to detect synthetic images based on Fourier-domain artifacts.
It supports training models with a fixed configuration, evaluating them on datasets,
and running inference on individual images or via a Streamlit demo.

---

📦 Installation

We use Astral uv for dependency management and execution.

Clone this repository and install dependencies:

uv sync

---

📂 Dataset structure

Datasets must be organized with the following structure:

dataset_root/
  train/
    real/...
    fake/...
  val/
    real/...
    fake/...
  test/
    real/...
    fake/...

• Synthetic test data: download from Zenodo
• Real test data: download raise-1k from RAISE

You can also have nested subfolders under real/ and fake/ (recursively scanned).

---

🏋️ Training a model (fixed config)

To train a model with a specific configuration:

1. Prepare a configuration file (JSON or joblib).
   Example config.json:

   {
     "dataset_root": "/path/to/dataset",
     "file_extension": ".png",
     "method": "rank",
     "rank_sz": 4,
     "max_period": 16,
     "jpeg_mode": "none",
     "learning_rate": 0.05,
     "max_iter": 200,
     "max_depth": 3,
     "max_leaf_nodes": 31,
     "min_samples_leaf": 50,
     "l2_regularization": 1e-6,
     "random_state": 753
   }

2. Train:

   uv run train_fixed.py \
     --config config.json \
     --save-model models/model.joblib \
     --save-config models/config.json

This will save both the trained model (.joblib) and the config used (.json).

---

🔍 Inference and evaluation

1. Single image inference

Run the trained model on a single image:

uv run infer_image.py \
  --model models/model.joblib \
  --config models/config.json \
  --image data/test/real/example.png

Outputs the predicted probability of being synthetic and the class decision.

---

2. Dataset evaluation

Evaluate the model on one of the dataset splits (train, val, test, or all):

uv run evaluate_dataset.py \
  --model models/model.joblib \
  --config models/config.json \
  --split test \
  --n-jobs -1 \
  --results results_test.json

Options:

• --optimize-threshold: re-optimize threshold for MCC on the chosen split.
• --force-jpeg-quality Q: compress all evaluated images at JPEG quality Q before evaluation.

Results are printed to stdout and saved in JSON (results_test.json).

---

3. Streamlit demo

A simple interactive demo is available:

uv run streamlit run streamlit_app.py

Features:

• Choose among all models in models/*.joblib
• Upload an image or provide an image URL
• See predictions live in the browser

---

Citation

@article{synthbuster,
  author={Bammey, Quentin},
  journal={IEEE Open Journal of Signal Processing}, 
  title={Synthbuster: Towards Detection of Diffusion Model Generated Images}, 
  year={2024},
  volume={5},
  number={},
  pages={1-9},
  doi={10.1109/OJSP.2023.3337714}}