Shulou(Shulou.com)11/24 Report--

With the fake and genuine AI generated pictures, AI can't tell them out by himself.

In this picture of Musk and his robot girlfriend, two of the five discriminators think it is true:

There is also this photo of a human and a 3m giant, which is judged to be true by five discriminators:

Ah, the AI discriminator doesn't seem to be very reliable.

This is a recent test conducted by the New York Times. They found five common AI discriminators on the market and fed them more than 100 photos for the test.

It turns out that the AI discriminator will not only mistake the AI photos for real, but also classify the real photos as generated by AI.

And the level gap between different discriminators is not small.

What's the specific performance? Let's take a look.

Adding some particles (Grain) can fool the discriminator. A total of five AI discriminators were used in this test, which are:

Umm-maybe

Illuminarty

A.I or Not

Hive

Sensity

The test included AI and human-created images, which were fed to each discriminator to see how they would judge.

AI authoring tools used include Midjourney, Stable Diffusion, Dall-e, and so on.

The New York Times mainly shows these examples. Contains 5 pictures created by AI, as well as 2 photos taken by real people.

From the statistical results, only Hive is correct among the five discriminators.

Umm-maybe performed worst, judging only two pictures correctly.

For example, this photo was generated by AI, and I heard that it won a big prize in a photography competition in February, which baffled most discriminators.

But this photo generated by pure AI has not escaped the eyes of most discriminators.

For human-created photos, the correct rate of AI discriminator is relatively high, and only Umm-maybe discriminator is wrong in both photos.

In addition, they specifically tested the art painting and found that most AI discriminators could tell that it was painted by a real person.

Compared with another painting created by AI, the same four discriminators are correct.

(Umm-maybe). It's really not very good)

It is worth mentioning that if some processing is performed on the AI image, the AI discriminator will fail.

For example, in this picture of Nike man, at first there are four discriminators that determine that it was generated by AI.

But if you add some particles to the image, the AI discriminator will judge the AI content of the image from 99% to 3.3%.

Finally, we also tested some discriminators (Umm-maybe, Illuminarty, A.I or Not) which can be used in real measurement.

It turns out that for Musk in the Soviet Union, Umm-maybe thinks there is an 85% chance that it is created by humans.

Illuminarty thinks there is only a 5.4% chance that it is created by AI.

Only A.I or Not determined that it was generated by AI.

What are the criteria for AI identification? So how on earth does AI tell the truth from the fake?

Generally speaking, they are different from human judgment criteria, which are generally based on the rationality of image content, while AI starts with image parameters, such as pixel arrangement, sharpness, contrast and so on.

So this explains the picture of the giant at the beginning, why all the discriminators feel so real.

More than a year after the AI painting fire, there are a lot of discriminators on the market.

Some are put directly on Hugging Face for free use, while others have set up companies that only provide API interfaces.

For example, Hive is a company that provides commercial solutions. From the test results above, we can see that the performance of Hive is also the best, and almost all of them can be judged correctly.

Before that, their main business was to provide data audit services for platform websites, supporting images, video and text, and serving platforms such as Reddit, Quora and so on.

Reference link:

Https://www.nytimes.com/interactive/2023/06/28/technology/ai-detection-midjourney-stable-diffusion-dalle.html

This article comes from the official account of Wechat: quantum bit (ID:QbitAI), by Mingmin.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.