🏠 Aksel

it works every time, sometimes

Detect AI photos using noise

#tech, #projects

Result: No photo submitted


So how does this work? #

Just to preface this, I'm not an expert and am simplfying a lot.

AI Image Generators start off with noise. The type often used is "Gaussian noise", which looks something like this.

Cloud-looking Gaussian Noise

A key defining point of this noise is that it has an average brightness of around 50%. Putting this image into photoshop and putting an average blur across the whole image:

Filter > Blur > Average in Photoshop

Picking the color, the brightness is almost always in the range from 40 - 60%.

The average brightness is 55%

Notice anything weird about these images?

The average brightness is 55%

These (obviously AI images) all have extremeley high contrast. This is due to the noise pattern! Many AI images have weird spots that are very dark because of this. In fact, taking the average brightness of all of these images:

The average brightness is 55%

We can see that the average brightness all sit in this 40 - 60% range. Of course, it is entirely possible to take a real photo with a brightness in this range, but for the most part, real images fall outside of this range.

More advanced AI generation programs are more advanced, often doing some post processing which makes it harder to detect. Check out this video from Corridor Digital for more manual ways to detect AI images (note: it's entirely possible to do it code, but ChatGPT only gets me so far.)

Try to notice the high contrast next time you look at an AI image, it's pretty obvious.

Thanks for reading!

Contact me: web@akselsalmi.com