Deepfake Detect

Deepfake Detect

A detector that tells real pictures from GAN-generated ones, using an ensemble of EfficientNet-B4 networks. It suits fact-checkers and researchers, with a Gradio page and a script.

Made by artspark on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Deepfake Detect is a Gradio app on Hugging Face Spaces that tells real pictures from ones made by a GAN. It combines several convolutional networks built on the EfficientNet-B4 backbone, each trained in a different way to hold up under compression and resizing. The repository also holds a script, an ROC notebook and the license file.

Who it suits: It suits people who need to check whether an image looks synthetic, and researchers who study image forensics.

Industry | Who uses it | How they use it | Example
Journalism | Fact-checkers | Check whether a suspicious picture may be generated | Run a picture through the detector
Trust and safety | Moderation teams | Screen uploaded images for synthetic ones | Score an image from the command line
Academia | Forensics researchers | Study the detector and its ROC curves | Open roc_curves.ipynb

Anyone who needs a guaranteed verdict can skip it, because a detector returns a score and the source gives no accuracy promise.

Where to use it:

• A web browser, straight from the Space page on Hugging Face
• Windows, macOS or Linux with Python, as a local copy
• A conda environment, for the detector script

Use case | Who | What they get back
Check one image online | Anyone curious | A score for whether the picture looks real or synthetic
Score images from a script | Researchers | The model score for a single image path
Review performance | Researchers | ROC curves for each dataset from the notebook

How to set it up: Using the Space online needs nothing. A local copy of the Space needs these first.

• Python and pip
• The huggingface_hub package, to fetch the files
• The packages in requirements.txt, plus Gradio
• For the detector script, a conda environment from environment.yml and the model weights, which the README says to download and unzip under the main folder

1. Install the Hugging Face Hub package.
2. Download the Space into a folder called deepfake-detect.
3. Move into that folder.
4. Install Gradio, then the packages from requirements.txt.
5. Start the app.

pip install -U huggingface_hub
hf download artspark/deepfake-detect --repo-type space --local-dir deepfake-detect
cd deepfake-detect
pip install gradio
pip install -r requirements.txt
python app.py

For the detector script on its own, create the conda environment, unzip the downloaded weights and run the script on one image.

conda env create -f environment.yml
conda activate gan-image-detection
unzip weigths.zip
python gan_vs_real_detector.py --img_path $PATH_TO_TEST_IMAGE

How to use it: Open the Space page on Hugging Face and give it an image. From a terminal, run the detector script with the path of one image.

Tool | What it does
app.py | Runs the Gradio page where you give an image
gan_vs_real_detector.py | Gives the model score for one image path
roc_curves.ipynb | Computes the ROC curve for each dataset

Example: You run the detector script on one photo and pass its path. The script gives back the model score for that picture, which tells you how synthetic it looks.

For newcomers and experts: If you are new, try the Space page with a photo you know is real. A GAN is a kind of program that generates pictures, and this tool scores how likely a picture is one of those.

If you know the field, the detector is an ensemble of EfficientNet-B4 networks trained with different recipes, following a published paper on training under JPEG compression. That aims to keep it working when images are compressed or resized.

You give the app an image
Several EfficientNet-B4 networks score it
The scores combine into one result

Good to know:

• Weights come as a separate download that the README points to, and the script needs them in the main folder
• A score is evidence, not proof, so treat it as one signal
• License: gpl-3
• Cost to run: none named
• Data it reaches: the images you give it and a weights download from the author

On Hugging Face, 13 November 2026

• Likes: 0
• Built with: gradio
• License: gpl-3.0

Links

• Space: https://huggingface.co/spaces/artspark/deepfake-detect
• Files: https://huggingface.co/spaces/artspark/deepfake-detect/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.