Applications like Google Photos' image search are a terrific example of how neural networks can do the great and useful job of recognizing what is in a picture and helping you search for them. But what about going in the other direction? It's a little iffier.
, some AIs with ginning up pictures based on text prompts, so far as the scope of the project is relatively constrained. If an AI only has to worry about birds, it can do alright at generating pictures. The same algorithm trained on a much wider variety of images, however, is significantly worse at its job and dishes out more surreal results. How much worse and how surreal? You can check it out yourself with put together by . Here's a taste of what it can (or can't) do.
Obviously there's a lot of room for improvement but maybe that's for the best. A tool like this that works perfectly is probably more horrifying than one that doesn't.