Erase the private parts. Keep the photo.

Faces, number plates, text, handwriting and codes, covered in solid black. The camera and location data stripped. A clean copy written from scratch, on your own machine, with nothing uploaded.

A parcel label showing a name, a street address and a phone number. — the photo as it was Drag the line. Invented name, address and phone number; the stripes are drawn, not a real barcode, so there was nothing for the decoder to read. the file it saved

The same label after ConsentGuard ran: the name, address, phone number, courier and order code are covered by solid black blocks. — the photo as it was Drag the line. Invented name, address and phone number; the stripes are drawn, not a real barcode, so there was nothing for the decoder to read. the file it saved

Get ConsentGuard 1.0 See how to run it Free and open source, and it stays on your machine.

Windows, PowerShell, a Python environment No account, no network calls, no telemetry Your working copy is deleted when you close the tab

Four frames, and the last one is yours to check

Real screens, running the real detectors on the parcel above. scroll the sheet

Add a photo

The upload area: a wide drop zone held by marker crop corners, with the two actions beneath it.

Drop it, paste it with Ctrl V, or pick a file. What it looks for is listed in the open, so you can switch any of it off.

It reads the photo

The scanning screen: the parcel photo with a marker-coloured line sweeping down it, beside the list of things being looked for.

Detectors cross the whole picture, then make a second, closer pass over small areas. Nothing leaves the machine while they run.

Check the cover

The editor: the parcel photo with hatched black blocks over every line of the label, and the tools above it.

Hatched black is exactly what gets saved. Paint over a miss, erase an over-reach, and use the loupe on the edges: B E L.

Look, then download

The result panel: ready to download, the reminder to look at the photo, and the checks run on the saved file.

The file is written fresh, re-opened and scanned again. Every check is listed in plain words, and you decide it is safe to share.

What it misses, measured and published

Most tools tell you what they catch. This is the whole scorecard from 120 validation photos the models never trained on, worst first, with the bad rows left in.

The bar and the number are the same measurement: how often a whole face, plate or signature was caught at all.

The line under each row says how much of that thing's area ends up covered when it is caught, which is usually much higher.

Medicine, documents 7% Weak. 52% of the area covered when caught. Check these by eye, every time.

Signatures 23% 66% of the area covered when caught.

Nudity 33% 93% of the area covered when caught.

Handwriting 50% 49% of the area covered when caught.

Number plates 56% 95% of the area covered when caught. The small and distant ones are what get missed.

Whole people 83% 95% of the area covered when caught. Off by default: turning it on erases most of a photo.

Faces 95% 99.8% of the area covered when caught.

QR codes, barcodes A decoder rather than a model, so it either reads the code or it does not.

Location, camera data Always removed. The file is written from scratch, then re-opened and checked.

It erases things you wanted to keep

All text goes, by design. On 30 validation photos with nothing private in them, every one had something erased, 45% of each photo on average. Switch text off when you only want faces and plates gone.

It reports its own false alarms

On the parcel photo above, the re-check said a number plate might still be visible in the saved file. There is no plate in that picture — the white label is plate-shaped. You get told, rather than quietly passed.

You are the last check

Detection is automatic and never complete, so the reminder to look at the photo sits next to the download button, and the app labels the content types it is bad at instead of pretending otherwise.

Run it on your own machine

Download release 1.0 and unzip it, or clone the same code:

git clone https://github.com/Nikhi00718/consent-guard.git
cd consent-guard

First time only, to build the Python environment and load the detectors:

powershell -ExecutionPolicy Bypass -File main_project\scripts\stage_02_baseline_model\setup_environment.ps1

Every time after that. The detectors load in about a minute, then it opens at http://127.0.0.1:7860:

powershell -ExecutionPolicy Bypass -File main_project\scripts\stage_05_review_export\start_consentguard.ps1

-ExecutionPolicy Bypass lets Windows run these two scripts from the folder you unzipped, without changing that setting for anything else. Both scripts are in the repository, and you can read them before you run them.

What it checks on the file it just saved

Opens as a normal imageRe-opened at full size without errors
Written as a new fileEncoded from scratch and unchanged since
No location or camera dataNone left in the file
Text, codes, faces, platesThe saved file is scanned again for each of them

Get ConsentGuard 1.0

Then look at every photo before you share it. That part is still yours.