🔬 UniversalFakeDetect: AI-Generated Image Detector

Towards Universal Fake Image Detectors that Generalize Across Generative Models (CVPR 2023)

Upload an individual image to evaluate whether it exhibits patterns characteristic of real photographs or synthetic AI generation.

⚠️ Limitation Notice: This detector identifies learned patterns associated with synthetic or manipulated images. It does not guarantee image authenticity.

Technical Notes:

  • Evaluates one image at a time.
  • Converts input to RGB and applies standard CLIP bicubic 224×224 normalization.
  • Features extracted using frozen OpenAI CLIP ViT-L/14 backbone with a trained linear probe.
Upload an image and click Analyze Image to begin.
  • Model: UniversalFakeDetect (CLIP:ViT-L/14 Backbone + Linear Classifier)
  • Paper: Towards Universal Fake Image Detectors that Generalize Across Generative Models (CVPR 2023)
  • Authors: Utkarsh Ojha*, Yuheng Li*, Yong Jae Lee (Wisconsin AI and Vision Lab)
  • Checkpoint: fc_weights.pth (trained with fixed CLIP backbone on ProGAN synthetic images)
  • Decision Rule: Sigmoid probability $\ge 0.50 \implies$ FAKE, $< 0.50 \implies$ REAL