Human Biometric Analysis

AI models learn hand appearance from training data without understanding that humans have exactly 5 fingers. They often generate 4, 6, or merged fingers because hands are complex and variable in photos.

MediaPipe detects 468 facial landmarks and measures the ratio between eye sizes, positions, and angles. Real faces have natural asymmetry within specific ranges, while AI often creates unnaturally symmetric or asymmetric eyes.

Yes, especially portraits with visible hands or multiple people. Even high-quality AI portraits often have subtle facial proportion errors or unnatural skin smoothness that this detector identifies.

The detector analyzes pore patterns, wrinkle consistency, and texture frequency. AI skin often lacks natural pore variation, has overly smooth areas, or shows repetitive texture patterns not found in real skin.

Biometric analysis is most effective on photorealistic images. Cartoon or stylized art intentionally deviates from human anatomy, so other detection methods are more suitable for those styles.

The system measures eye-to-nose ratio, nose-to-mouth distance, face width proportions, and ear placement. AI often generates faces that look good at first glance but have proportions outside natural human ranges.

MediaPipe provides the best accuracy, but the system includes fallback methods using OpenCV Haar cascades when MediaPipe is unavailable. The fallback has reduced accuracy but still detects major anomalies.

DALL-E 3 has significantly improved hand generation, but still makes errors in complex poses or multiple hands. The detector catches these remaining issues, though accuracy is lower than with older models.

If no humans, hands, or faces are detected, this method returns a neutral score (0.5) and other detection methods carry more weight. The ensemble system automatically adjusts for content type.

MediaPipe is optimized for real-time processing. Full biometric analysis typically completes in 200-500ms on CPU, making it suitable for production use without significant latency.