The moment you upload, two things happen invisibly. First, your image is treated with a layer of adversarial protection — microscopic changes imperceptible to the human eye that cause AI training models to misread your work. If someone tries to train an AI on your images, it learns the wrong thing.

Glaze/Nightshade (Adversarial Poisoning): * How: Use the Glaze or Nightshade CLI/API (developed by University of Chicago). It applies microscopic pixel shifts that trick AI training models into misinterpreting the image's style or content.

Second, your image receives a visual fingerprint — a unique signature based on what the image actually looks like, not just its file data. Unlike a standard checksum, this fingerprint survives cropping, screenshotting, color adjustments, and watermark removal. It's how we know it's your image even when someone has tried to hide that fact.

DINOv2 Perceptual DNA (Fingerprinting):* How: Use Meta’s DINOv2 model to generate a 384-dimensional vector.

Why DINOv2? Unlike standard pHash, DINOv2 understands features. If someone screenshots the image, crops out the watermark, and flips it, the vector remains 98% identical.

Storage: requires vector database like Pinecone or Qdrant for instant "Reverse DNA" lookups.