Documentation & Mechanics
AI models "see" images as mathematical tensors (grids of numbers). We inject a calculated "digital mist" (Gaussian noise optimized via gradient descent) that aligns specifically with the features neural networks look for. To a human, the image is unchanged. To an AI, the feature vectors are scrambled.
Uses Fast Gradient Sign Method (FGSM) on VAE Latent Space.
Uses deep iterative PGD attacks targeting the exact CLIP/UNet visual encoding layers.
Processing occurs entirely in Volatile Memory (RAM). No data is written to disk. The session destroys all tensors immediately after the result is returned.