Scope & Limitations

What Facial Authenticity Localization detects, what it misses, and why context matters.

Explicit Boundaries False Positives/Negatives Non-Forensic

1. Detection Scope vs. Non-Scope

Face Warp Check is purpose-built for a specific subset of digital photo editing: geometric pixel warping. It does not attempt to solve generalized image authenticity.

In Scope (Targeted) Out of Scope (Not Evaluated)
Photoshop Face-Aware Liquify Diffusion / GAN deepfakes (Midjourney, DALL-E)
Facetune / Meitu reshaping brushes Skin smoothing / blemish removal
Manual warp / pinch / expand transformations Color grading, filters, and white balance
Facial symmetry drag distortions Lighting tricks, angles, makeup, posture changes

2. Key Failure Modes & Constraints

Critical Technical Constraints

  • Requires a visible, unoccluded face: The model requires visible facial anatomy (eyes, nose, mouth) to compute normalized landmark bounding boxes. Extreme profiles or heavy sunglasses will cause alignment failure.
  • Vulnerability to heavy compression: Aggressive JPEG compression (e.g., screenshots of Instagram posts re-saved multiple times) disrupts the sub-pixel interpolation artifacts that FAL searches for, degrading detection accuracy.
  • Lens distortion edge cases: Wide-angle smartphone cameras held close to the face introduce barrel distortion. While distinct from Liquify warping, extreme perspective shifts can occasionally create noise in flow maps.

3. Epistemological & Forensic Rules

We strictly adhere to scientific reporting standards to prevent misinterpretation:

  1. A low score is not proof of honesty: An image with a low score may still feature physical deception, heavy makeup, dramatic lighting, or AI-generated inpainting that bypassed geometric warp classification.
  2. A high score is not proof of identity or culpability: A high score indicates warping signatures in the digital file. It does not establish who performed the modification (e.g., patient, photographer, agency, social media app automated preset).
  3. Signal, never a binary verdict: Digital forensics operates on probabilities and signature matches, never deterministic absolute truth.