Humans struggle to distinguish between real and AI-generated faces

What’s real?

Scientists use sophisticated algorithms to generate faces with the same bone structure and skin density as pictures of real people. And we like what they’ve created.

But how will you know if someone is real or not?

According to the article https://www.pnas.org/doi/10.1073/pnas.2120481119

This is because AI-generated faces have become so advanced that humans now cannot distinguish between real and fake.

 

Generative adversarial networks (GANs) are popular mechanisms for synthesizing content and I’m using it myself for long time now for deepfake and GAN image generation, to say the truth, the faces its generating looks very convincing, But a sharp eye will tell the difference in no time. The problem is that in most of the cases the watcher will not look for the imperfection that AI can create and will believe the images are real. From my perspective Its good that the AI can generate such a realistic looking face as we have a real problem of AI bias that is a result of small amount of training data that can be fixed by creating the missing information by GAN and other similar engines.

Here are the most (top and upper-middle lines) and least (bottom and lower-middle) accurately classified real (R) and synthetic (S) faces:

From the paper its clear that a better machine needs to be developed in order to check other machines

Some kind of VQGAN vs CLIP that work as a pair, one is painting and the other making corrections.

But in the same time, I really think that we need tools that can produce something that is hard to distinct as fake or not in order to create a better trained engines with less bias.

Add a Comment

Your email address will not be published.