The true pixel

Using statistics to prove the level of confidence | CODE

How to detect fake pictures made by AI

Most generative tasks, notably image-to-image tasks like super-resolution and inpainting, use diffusion models. However, due to uncertainty, diffusion-based approaches are rarely used in high-stakes procedures.

For example, if doctors look at a low-resolution MRI scan, increasing its resolution may greatly help them. But because generative models tend to "hallucinate" details, the doctor might not be able to trust that a high-resolution image is true. By giving the doctors a statistically certain interval around each pixel, the doctors can use what the generative model comes up with.

Trust in Diffusion Models

The following method allows you to determine the "true" of a pixel: