Blur face areas in a photo

Blur face areas only after reviewing every suggestion

Press Suggest face areas only after the image is visible. A self-hosted BlazeFace model returns rectangles around likely faces, and the page expands them to include a margin around hair and chin. Those rectangles are proposals. Remove false matches, inspect the background at full size, and drag extra boxes over every person the detector missed.

Small faces, profiles, masks, motion blur, screens inside the photo and partial occlusion are ordinary failure cases. A product that exported immediately after detection would turn a probabilistic model into an invisible privacy promise. This one keeps review in the middle and labels each rectangle as detector or manual.

Strong pixelation and strong blur are still concealment

Pixelation reduces each face area to large color blocks. Blur applies a strong radius inside the reviewed rectangle. Both effects are flattened into a new PNG or JPEG rather than saved as removable layers. They make a face difficult to recognize casually, but they do not provide the same guarantee as deleting the area.

For identity documents, account numbers, private text or any case where recovery risk matters, use solid image redaction. It replaces the selected pixels with opaque black. Face blur is useful when the visual context should remain while people become less identifiable; it should not be marketed as cryptographic erasure.

The local detector loads only after your click

The page shell does not carry a computer-vision model. The first detection action fetches the versioned MediaPipe Tasks Vision WebAssembly runtime and a 224 KB BlazeFace short-range model from CreatorValet's own origin. The model and runtime are Apache-2.0 assets stored beside their license and exact source information. They can be cached by the browser.

Inference runs against the local preview canvas. No photo bytes, embeddings, face landmarks or rectangles go to Google, a CDN or an API. The model recognizes face-shaped regions; it does not identify who a person is and performs no face recognition or database lookup.

Why manual additions remain available

Drag on the image to add a box at any time, even if the detector failed to load. Manual boxes also handle reflections, faces on posters, license plates and other contextual details you want to conceal. The rectangle list lets you remove one proposal without rerunning the model.

Detection starts a new proposal set so stale boxes from a different model run do not hide under the new receipt. Before export, count the people in the full photo and compare that number with the reviewed regions rather than trusting the detector count alone.

PNG and JPEG create new flattened images

PNG preserves screenshot and graphic edges without loss. JPEG usually makes photographs smaller but recompresses visible detail. Both downloads are encoded from a browser canvas, so the original EXIF, GPS, XMP, thumbnail and provenance blocks are not copied. If metadata removal is the only job, the EXIF remover gives a more targeted report.

Open the downloaded file before sharing. Zoom into each rectangle, look for a missed face at the edges or in the background, and confirm that the target platform receives this new file rather than an original cloud-library version. Use the image compressor only after that review if a size limit remains.

Short-range means photographs, not surveillance

BlazeFace short-range is designed for relatively close faces and fast local detection. It is a practical match for portraits, group photos and phone pictures, not a guarantee for a stadium crowd or tiny distant people. For dense scenes, manual inspection and boxes are mandatory.

The tool processes one still image. It does not track faces through video, identify the same person across files or produce biometric templates. Those would be materially different products and privacy boundaries.

Verify that processing stays local

The local-processing guide shows how to open the Network panel before detection. You should see versioned runtime and model files coming from CreatorValet, but never the chosen image in a request body. Reloading clears the image and reviewed boxes from memory.