Your feed is full of AI-generated faces pretending to be real people. Perfect skin, flawless lighting, zero pores — that's not beauty goals, that's a diffusion model. Drop any photo below and we'll run 12 forensic scans to tell you the truth.
🔬 FFT Spectral Analysis
🧠 Claude Vision AI
🔍 Error Level Analysis
🧬 Skin Pore Detection
How it works: Upload → we run two-stage AI forensic analysis (ForenX method, CVPR 2025) + 10 client-side algorithms including cross-difference filtering, frequency band deviation, and patch-based texture fingerprinting. Results in seconds.
Protect Your Image
AI bots are actively scraping social media to harvest real photos — then using them to train models that generate fake influencers, deepfakes, and synthetic profiles built from your likeness. Don't wait until your face shows up on a fake account. SafeFace adds invisible protection to your photos before you post — defend your identity now.
🧠 Claude Vision AI — Two-Stage Forensic Analysis
Error Level Analysis (ELA)
Recompressed at Q=8%. Uniform ELA = likely AI; varied levels = real photo with natural compression history.
FFT Frequency Spectrum
Based on SPAI (CVPR 2025): AI images show spectral artifacts. Real images have natural high-frequency roll-off; AI images show anomalous peaks or uniformity.
Cross-difference residual
Forensic Signal Breakdown
Methodology & References
This tool implements techniques from:
ForenX (2025) — forensic prompting for MLLMs;
SPAI (CVPR 2025) — spectral analysis for AI detection;
DEFEND (ICLR 2025) — frequency band deviation;
GPT4V Forensics (ACM 2025) — two-stage LLM prompting achieving 92.1% accuracy; Multi-query averaging per CVPR 2024 workshop findings; Cross-difference filtering from Synthbuster (Bammey 2024).
Disclaimer: Results are probabilistic. No AI detector achieves 100% accuracy. This tool combines multiple research-backed approaches for the best possible client-side + LLM assessment.