Home Services About Blog Tools Contacts
Start an engagement Emergency call
Englishen Españoles Magyarhu العربيةar Русскийru Українськаuk Deutschde Slovenčinask ไทยth 中文zh-CN 日本語ja 한국어ko Românăro Françaisfr हिन्दीhi বাংলাbn Bahasa Indonesiaid Portuguêspt Italianoit Tagalogtl Tiếng Việtvi فارسیfa Kiswahilisw မြန်မာmy Türkçetr اردوur Basa Jawajv Polskipl Češtinacs Afrikaansaf తెలుగుte

Detect AI-Generated Photo

Select photo to test

Heuristic analysis runs entirely in your browser — no data leaves your device.

Click to select or drag & drop

JPG and PNG — analyzed locally

Selected image for analysis
Technical Details
AI Evidence Real-Photo Evidence
Noise Residual --
Noise residual heatmap
ELA Map --
Error level analysis map
Anomaly Heatmap --
Residual anomaly heatmap
Saturation Map --
Color saturation heatmap
Frequency Map --
High-frequency detail map
Gradient Directions 18 bins
Luminance Histogram 64 bins
Heuristic signals

    Results are evidence scores, not proof of origin. The model is optimized to prefer UNCERTAIN over false accusations when evidence is weak.

    • Format conversion alters noise and frequency signatures and may increase false-positive rate.
    • Messenger compression (WhatsApp, Telegram, etc.) re-encodes and strips metadata — photos saved from chats may show AI evidence even when real.
    Model Card v5.9.0
    Decision bands
    AI evidence >= 64%, real-photo evidence <= 55%, otherwise UNCERTAIN.
    Runtime
    Browser-only. A deep EfficientNet-B0 classifier (ONNX, WASM) is the primary signal, weighted above the heuristic + k-NN mini-ML and safety caps. If the model fails to load, the detector falls back to the heuristic + k-NN ensemble.
    Limits
    Small, heavily compressed, screenshot-like and converted files should be treated as low evidence.
    Need more than a free tool?

    PWN-ALL is a licensed cybersecurity firm — this tool is built by the same team that runs real-world penetration tests against infrastructure, web apps and APIs.

    Request a penetration test