An explanation of inception scores

In this video, former TechTarget editorial assistant Melanne Ghahraman talks about inception scores.

We've all had fun asking AI platforms to generate images -- some more successful than others. But aside from human judgment, there's a way to score how accurate an AI image is to the prompt it was given.

It's called the inception score -- or IS -- a mathematical equation that's just too complex for us to get into right now. But basically, it scores images on a scale from zero (being the worst) to -- theoretically -- infinity (the best).

The IS is based on two factors: quality, or how good the image is; and diversity, or the variety and randomness of the images it produces.

For instance, say the AI is producing images of cats. Quality images should have a clearly identifiable cat, like if you'd drawn a picture or taken a photo of a cat. If the image is not clearly identifiable as a cat, the IS will be low.

Diversity would mean producing a different cat breed or cat pose every time. If the AI produces the same kind of cat over and over, the IS will be low.

There are some limitations of the algorithm, though. It only works on small, square image sizes; a limited sample size will produce an artificially high score; and unusual images that weren't part of the training set will produce an artificially low score.

Sabrina Polin is a managing editor of video content for the Learning Content team. She plans and develops video content for TechTarget's editorial YouTube channel, Eye on Tech. Previously, Sabrina was a reporter for the Products Content team.