Toggle Dark Mode
Want to know when someone edited a selfie to make themselves look more attractive? Do you wonder if they altered their nose or the color of their eyes? If you’ve answered yes, then we have the tool for you.
As posted on Reddit, the new Mirage app from Akshat Jagga does just that. It detects when an image has been edited and tries to identify those edits. When it’s successful, you’ll get a glimpse of the original photo (for better or for worse).
The powerhouse of the app is machine learning that can detect when an image has been edited.
This functionality is based onÂ codeÂ developed by researchers from Adobe Research and UC Berkeley. The team publishedÂ a research paperÂ in 2019 that describes a method of detecting facial warping manipulations in an image.
The tool was so good at detecting deception that it even outperformed human examiners who were asked to look closely at the pictures for signs of manipulation.
The researchers’ tool, and the Mirage app, not only can tell when an image was changed, but it also can use machine learning to figure out how the photos were altered.
It can predict which areas were manipulated and then tries to reverse these changes one by one. Once the algorithm is done scanning and identifying modifications, the tool then can show which sections in the photo were edited. It even reveals its best guess at the original image.
The app has one major flaw — it only works when a face is in the picture. You can’t use it to see if a car, tree or some other object was edited inside of a photo. It also is challenging to understand which sections were changed because they are often too small to see easily. The app is a work in progress, though, and the developer said he is working on an update that makes it easier to see the edited parts. Expanding the functionality beyond faces hopefully is also in the works.
The Mirage app is available now in the iOS App Store for $1.99. There’s also an Android version that is available in Google Play Store and requires a $0.99 monthly subscription.