🔍 Deepfake Detection Space

This space collects a series of state-of-the-art methods for deepfake detection, allowing for free and unlimited use.

Training & Performance

All methods have been trained using the DeepShield dataset, on images generated with Stable Diffusion XL and StyleGAN 2. You can expect performance comparable to the results shown in Dell'Anna et al. (2025).

Understanding the Results

  • Prediction: Tells if an image is Real or Fake.
  • Confidence: The confidence with which the model determines if the image is real or fake.
  • Elapsed Time: The time the model needed to make the prediction (excluding preprocessing or model building).

Understanding the Results produced by "ALL"

  • Runs all available detectors (R50_TF, R50_nodown, CLIP-D, P2G, NPR) sequentially on the input image.
  • Produces a Weighted Majority Vote verdict (Real/Fake). Each model's vote is weighted by a fixed importance score (summing to 1) based on user ranking and its confidence score. Only confident predictions (> 0.6) are counted.
  • You can find the specific weights used for each model in the "⚖️ Weight Details" menu below.
  • Also generates a Confidence Plot visualizing each model's score and a textual Explanation of the consensus.
  • In the plot, Green bars indicate a Real prediction, while Red bars indicate a Fake prediction.

Note

⚠️ Due to file size limitations, model weights need to be downloaded automatically on first use. This may take a few moments.
⚠️ To provide a free service, all models run on CPU. The detection process may take a few seconds, depending on the image size and the selected detector.
Select Detector

Choose which deepfake detection model to use

Detector Weights

The weights are assigned based on the ranking (based on the results of TrueFake: A Real World Case Dataset of Last Generation Fake Images also Shared on Social Networks): CLIP-D > R50_TF > R50_nodown > P2G > NPR, such that their sum equals 1.

Detector Weight
CLIP-D 0.30
R50_TF 0.25
R50_nodown 0.20
P2G 0.15
NPR 0.10

ALL

  • Description: Runs all available detectors (R50_TF, R50_nodown, CLIP-D, P2G, NPR) sequentially on the input image.
  • Results: Produces a Majority Vote verdict (Real/Fake) considering only confident predictions (> 0.6). Also generates a Confidence Plot visualizing each model's score and a textual Explanation of the consensus.

R50_TF

R50_nodown

CLIP-D

P2G (Prompt2Guard)

NPR

References

  1. Dell'Anna, S., Montibeller, A., & Boato, G. (2025). TrueFake: A Real World Case Dataset of Last Generation Fake Images also Shared on Social Networks. arXiv preprint arXiv:2504.20658.
  2. Corvi, R., et al. (2023). On the detection of synthetic images generated by diffusion models. ICASSP.
  3. Cozzolino, D., et al. (2023). Raising the Bar of AI-generated Image Detection with CLIP. CVPRW.
  4. Laiti, F., et al. (2024). Conditioned Prompt-Optimization for Continual Deepfake Detection. arXiv preprint arXiv:2407.21554.
  5. Tan, C., et al. (2024). Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. CVPR.