HAZE is for anyone who posts photos of themselves and does not want those photos fed to face-recognition AI. Journalists publishing photos of sources. Activists publishing photos of protests. Survivors of stalking. Parents posting photos of kids. People who simply do not want their face indexed by every AI company that scrapes the web.
HAZE loads your photo on your computer. It runs a small face-detection model to find each face in the photo. Then it adds a carefully designed pattern of tiny color changes on top of the face area. The changes are too small to see (a few units out of 255 on each color channel) but they are designed to land in the frequency range that face-recognition models pay most attention to. The result: a photo that looks identical to you but that face-recognition systems struggle to match against other photos of the same person.
HAZE does not have weak modes. Every cloak runs at the strongest setting. Giving users a "low strength" option would let them ship photos with less protection by mistake. HAZE protects every face every time, as much as the math allows.
HAZE makes face recognition harder. It does not make face recognition impossible. We did not train HAZE against any one specific face-recognition system. Results vary by which system is trying to identify you. New face-recognition models can sometimes defeat older cloaks. HAZE is one layer of privacy, not the only one. The app states this plainly when you first open it.
HAZE is also not a deepfake detector, not a face blur, and not a fake-photo generator. It just adds invisible patterns to existing photos of real faces.
Full cloak at maximum strength. Cloak one photo at a time. Before and after slider. Save the cloaked photo wherever you want. Auto-updating face-detection model. Strips GPS location from the saved file.
Everything in Free. Plus a PDF processing receipt for newsroom accountability. Cloak many photos at once. Pick which faces to cloak in a group photo (e.g. the child only, not the journalist).
HAZE makes zero outbound network calls. Your image is loaded on your computer, processed on your computer, and saved on your computer. Cinderpoint never sees it. The face-detection model is bundled inside the app, and the App Store handles app updates including any refreshed model. Advanced users can download an alternative model from the UltraFace open-source project on GitHub and import it through Settings.
Read the privacy policy for the full statement.
HAZE is built on Electron, React, TypeScript, and ONNX Runtime. The face-detection model is UltraFace, an MIT-licensed open-source model from the Linzaer project. The perturbation pattern is a face-localized high-frequency adversarial pattern, generated deterministically per image. All bundled open-source dependencies are credited in the NOTICE file shipped inside the app.
On research: HAZE is informed by the published Fawkes research (Shan et al., University of Chicago, 2020) which demonstrated that face-recognition systems can be defeated by image perturbations at the embedding layer. HAZE does not bundle a Fawkes-derived model and is not affiliated with the Fawkes project. The current HAZE perturbation attacks the general feature space, not any one specific recognition system.
Version 1.0 is built and tested. Store submissions in progress. Sign up below to be notified when HAZE goes live in your store.