A perceptual hash for the AI era

Classical perceptual hashes (pHash, dHash, aHash) were designed to survive JPEG compression and minor rescaling. They were not designed to survive a diffusion-based upscaler that rewrites every pixel while preserving semantic content.

Our hash, internally called PHash-V, operates on a learned embedding rather than raw pixels. It is designed to collide on semantic content (same scene, different generation seed) and not collide on cosmetic edits (crop, colour-grade, watermark removal).