Blur & Unblur AI is a focused browser tool for hiding faces before an image is shared. It loads JPG, PNG or WebP files into an HTML Canvas, detects faces locally when the browser or bundled fallback supports it, creates editable masks, lets the user remove false detections or draw missed regions with a lasso, and exports a PNG. No account or processing-server upload is used by the current implementation.
Despite the name, this is primarily a face-blurring privacy editor. Its “Face Unblur” mode does not reconstruct missing pixels or reverse a previous blur. Inspection of the published client script shows that it redraws the original selected region with modest contrast, saturation, brightness and soft-light enhancement. It can make an already visible face look punchier; it cannot recover ground-truth identity or text that was destroyed by blur.
What each mode actually does
| Mode | Published implementation | Use it for | Do not assume |
|---|---|---|---|
| Face Blur | Clips Canvas blur to enabled face boxes or lasso paths | Quickly reduce face visibility while preserving the rest of a photo | Blur is irreversible, legally sufficient or safe against determined recovery |
| Face Unblur | Applies contrast, saturation, brightness and soft-light enhancement inside selected masks | Cosmetic emphasis of an already visible face | Deblurring, deconvolution, super-resolution or factual detail recovery |
| Full Blur | Applies a Canvas blur filter to the entire image | Obscure a background or create a soft visual treatment | All sensitive metadata or recognizable context has been removed |
| Select | Toggle or remove detected masks | Correct false positives and choose which faces are affected | A removed mask was a duplicate rather than a real person |
| Lasso | Create a freehand clipping path | Add missed faces or cover non-face identifiers | A loose hand-drawn boundary covers hair, profile and surrounding identifiers |
How local processing works
The current public page and script decode the selected file with createImageBitmap, scale the working canvas to a maximum width of about 980 pixels, draw the source locally and export with canvas.toDataURL("image/png"). Face detection first tries the browser's native FaceDetector. If unavailable, it loads bundled tracking.js and face-classifier assets from the same site. If both fail, lasso mode remains available.
This supports the narrow claim that the source image is not posted to a processing endpoint in the inspected version. “Runs locally” is not the same as “the page makes no network requests”: the site still loads its application assets and a Google Fonts stylesheet, and a future deployment could change. For highly sensitive work, disconnect after the application loads, inspect developer-tool network traffic, use an offline editor or self-host reviewed code.
A safe face-blurring workflow
- Work from a copy. Preserve the original only in the authorized location; do not overwrite evidence or the sole archive.
- Remove metadata first. GPS, device, capture time, author and thumbnail metadata can identify people even when pixels are blurred.
- Load the copy and inspect detection. Zoom through the entire frame. Check profiles, reflections, screens, posters, children, distant people and partially covered faces.
- Correct every mask. Remove only confirmed false positives. Use lasso for missed faces and expand beyond eyes to cover the whole identifying region.
- Choose the redaction method by risk. Blur is suitable for casual sharing; use opaque replacement or strong pixelation/cropping for higher-risk anonymity.
- Export and reopen the file. Inspect the downloaded PNG at 100–400% zoom rather than trusting the editor preview.
- Test outside the app. Check thumbnails, messaging previews, OCR, reverse-image search and whether context still identifies the subject.
- Share the derivative only. Confirm the attachment is the exported copy, not the original with reversible editor state.
Why blur strength is not a privacy guarantee
Blur replaces each output pixel with a mixture of nearby source pixels. If the radius is weak, the original structure still influences the result. Depending on the blur, resolution and adversary, sharpening or deconvolution may reveal some information. More importantly, a face can remain recognizable from silhouette, hair, clothing, tattoos, location, companions and surrounding text even when facial features are softened.
For school images, abuse survivors, witnesses, health contexts, protests or legal evidence, use a destructive method designed for redaction: crop the subject out, cover the full identifying region with an opaque solid shape, flatten the image, remove metadata, then verify the exported pixels. Never place a semitransparent shape or an editable layer over the original and assume the hidden pixels are gone.
| Method | Visual utility | Privacy strength | Best fit |
|---|---|---|---|
| Light Gaussian blur | Natural-looking context | Low; identity cues often remain | Aesthetic softening, not sensitive redaction |
| Heavy blur | Context remains readable | Medium and source-dependent | Casual public sharing after review |
| Strong pixelation | Obvious edited region | Variable; block size must overwhelm face detail | News/social images with visible disclosure |
| Opaque replacement | Region is fully hidden | High for covered pixels after flattening | Sensitive identity protection |
| Crop/remove subject | Loses scene information | High when all identifying context is removed | Highest-risk publication |
| Synthetic replacement | Can preserve scene appearance | Complex; generated identity and disclosure issues | Specialist workflows with documented review |
Face detection is an assistive step
Native FaceDetector support is not uniform across browsers, and automated classifiers are sensitive to pose, scale, occlusion, lighting, image quality and demographic representation. The tool's source caps native detections at 20 faces. A crowded event photo may contain more people, so a nonempty result does not prove completeness.
The fallback can time out, and the canvas is downscaled for editing. Small distant faces may become harder to detect. A privacy workflow therefore needs a human completeness review, not merely a “faces found” count. For group photos, make a checklist that moves systematically from top left to bottom right and includes mirrors, windows and background displays.
Input, preview and output behavior
| Area | Current behavior | Practical consequence |
|---|---|---|
| Input | JPG, PNG and WebP accepted by the file picker | RAW, HEIC, PDF and animated images need conversion or another tool |
| Working size | Canvas is limited to roughly 980px wide; smaller images remain at original size | Export can be lower resolution than a large source photo |
| Detection | Native FaceDetector, then local tracking.js fallback, then manual lasso | Behavior varies by browser; lasso is the reliable fallback |
| Blur strength | Slider maps to a CSS/Canvas blur radius with a minimum effect | Percentage is not a standardized privacy score |
| Export | Flattened PNG generated from the working canvas | Guide boxes are omitted, but JPEG source may become a larger PNG |
| History | No durable project or undo stack is documented | Complex correction may need restarting from the source copy |
Evaluation test set
Before relying on the tool, test at least 30 representative images: frontal and profile faces; very light and dark scenes; different skin tones and ages; glasses, masks and hats; partial faces; large crowds; mirrors; printed photographs; faces on laptop/phone screens; and high-resolution camera files. Record missed-face rate, false-positive rate, time to correct masks, final export resolution and whether any reviewer can still identify the protected subject.
Alternatives
| Option | Prefer it when | Tradeoff |
|---|---|---|
| Blur & Unblur AI | A fast, free, local browser workflow is enough | Limited formats, resolution, history and verification tooling |
| Photoshop or Affinity Photo | You need precise masks, layers, original resolution and repeatable export | Paid software and a longer learning path |
| GIMP | You need a free offline editor with destructive opaque redaction | More manual steps than a focused face tool |
| ImageMagick | Batch processing and reproducible command-line redaction are required | Face masks must come from another detector or coordinates |
| Mobile photo editor | One image must be handled quickly on a phone | Cloud upload, metadata and export behavior vary widely |
| Dedicated evidence-redaction system | Legal, healthcare, public-record or investigative auditability matters | Higher cost and process overhead, stronger controls |
Frequently asked questions
Does Face Unblur recover a blurred face?
No. The inspected implementation applies cosmetic contrast, saturation, brightness and soft-light changes to the original selected pixels. It does not reconstruct destroyed detail.
Are images uploaded?
The current tool decodes, edits and exports the source with browser APIs and has no image-processing upload endpoint in the inspected script. The page itself still loads web resources; verify the network behavior for sensitive use.
Is face blur enough to anonymize someone?
Not always. Weak blur and surrounding cues can preserve identity. Use opaque flattened redaction or removal when harm from re-identification is significant.
What happens when detection misses a face?
Use Lasso mode to draw a mask. Always review the complete image; automatic detection is a convenience, not a completeness guarantee.
Does export preserve the original resolution?
The source shows a working-canvas maximum width around 980 pixels, so a large source can be downscaled in the exported PNG. Check output dimensions before publication.
Does PNG export remove all metadata?
Canvas-generated PNG normally creates new pixel output rather than copying the original container, but do not treat that as a universal metadata guarantee. Inspect the actual downloaded file with a metadata tool.
Sources and verification
- Blur & Unblur AI official application
- Official privacy notice
- Published client implementation inspected for this review
- MDN FaceDetector API status and compatibility context
- MDN Canvas 2D filter reference
- MDN Canvas export reference
- NIST Privacy Framework for risk-based privacy controls
Last reviewed July 25, 2026 against the deployed page and published client bundle. A redeploy can change code or privacy behavior; recheck before handling sensitive images.
