µFlow — Leveraging Average Images for Improving Generalisation of Deepfake Faces Detectors

One-class deepfake face detector trained only on real images.  ·  Accepted at ECCV 2026.

µFlow is a one-class deepfake detector trained only on real images. It models the distribution of features extracted from average real images with a Gaussian Mixture Model (GMM), and trains a normalizing flow (FastFlow) to map the features of single real images into that distribution. At inference, the negative log-likelihood of an image is used directly as a fakeness score.

This repository documents how to train and test the model. For the method description, experiments and results, see the project page.

---

Repository layout

MuFlow/
├── config.py                   # dataset, families & reporting configuration
├── main.py                     # training entry point
├── eval.py                     # evaluation (robustness attacks & custom dirs)
├── configs/                    # one YAML per model config
├── data/                       # split CSV lives here
├── scripts/                    # dataset & GMM preparation
└── src/muflow/                 # the reusable library (model & method only)
    ├── model.py                # FastFlow model & backbone builder
    ├── dataset.py              # data loading & transforms
    ├── patch_utils.py          # native-patch sampling & feature extraction
    ├── calibration.py          # threshold calibration & sweep
    ├── attacks.py              # content-preserving degradations
    ├── constants.py            # backbone & normalisation specs
    └── gpu_utils.py            # automatic GPU selection

---

Installation

Requires Python ≥ 3.12 and a CUDA-capable GPU.

git clone https://github.com/opontorno/MuFlow.git
cd MuFlow

conda create -n muflow python=3.12 -y
conda activate muflow

pip install -e .

---

Configuring the datasets

All dataset and reporting settings live in config.py at the repo root — adapt it to your data without touching the library.

MUFLOW_DATA_DIR is required: config.py raises immediately if it isn't set (either export it, or hardcode DATA_DIR directly in config.py):

export MUFLOW_DATA_DIR=/path/to/your/data

The core of it is three glob patterns, shown below with the values this project uses for its own experiments — edit them to match your own folders and file names (there is no mandatory folder structure; a glob that matches nothing raises a clear error naming the offending pattern):

PATH_REAL     = [f"{DATA_DIR}/datasets/ffhq/**/*.*g"]        # train / val / calibration
PATH_REAL_OOD = [f"{DATA_DIR}/datasets/celeba_hq/**/*.*g"]   # test baseline (None → reuse PATH_REAL)
PATH_FAKE     = [f"{DATA_DIR}/datasets/WILD/**/*.*g", ...]    # fake generators (test only)

config.py also holds the generator families (for the per-family report), the console reporting flags, and the dataset-prep knobs (mean size, per-class cap, split ratios) — adjust the family names to match your own fake-generator folders.

For reference, the layout this project uses is the following — but any other arrangement works just as well, as long as config.py points at it:

$MUFLOW_DATA_DIR/
├── datasets/
│   ├── ffhq/<sub>/*.png                        # real — training source
│   ├── celeba_hq/{train,val}/<sub>/*.jpg       # real — test-time baseline
│   ├── WILD/{Closed_Set,Open_Set}/<gen>/*.png  # fake generators
│   └── other_sources/<gen>/*.png               # fake generators
└── datasets_means/<mean_size>/<source>/*.png   # average images

MuFlow/data/dataset_split_rand.csv              # train/val/test split

---

Training

Step 0 — Build the split CSV

The dataset is indexed by a CSV (data/dataset_split_rand.csv, columns path,split):

python scripts/generate_csv.py

Step 1 — Compute average images

Compute the average images of the real source (taken from PATH_REAL by default):

python scripts/generate_means.py

Pass --input <glob> --name <folder> to average a different source.

Step 2 — Fit the GMM

python scripts/generate_parameters.py

This step is optional: main.py runs it automatically if the parameters file is missing.

Step 3 — Train

python main.py

Each run writes a checkpoint (best.pt), the calibrated thresholds, predictions and a metrics JSON to $MUFLOW_CHECKPOINT_DIR/<run_name>/.

---

Testing

python eval.py --run_dir logs/<run_name>

Settings are loaded from the run's run_config.yaml. You can also evaluate on custom folders:

python eval.py --run_dir logs/<run> \
    --custom_dirs /path/reals /path/fakes --custom_labels 0 1

---

Citation

If you find this work useful, please consider citing:

@inproceedings{pontorno2026mu,
  title={{{$\mu$Flow}: Leveraging Average Images for Improving Generalisation of Deepfake Faces Detectors}},
  author={Pontorno, Orazio and Litrico, Mattia and Guarnera, Luca and Giuffrida, Mario Valerio and Battiato, Sebastiano},
  booktitle={European Conference on Computer Vision},
  year={2026},
  organization={Springer}
}

---

Contact

Orazio Pontorno — University of Catania — orazio.pontorno@phd.unict.it

---

License

See LICENSE.