Is It Real, Or Is It An AI Generated Image
Your computer is now officially an AI, at least when it comes to generating realistic images from just about any source. With the previous so called AI image generators like DALL-E 2 and Imagen were cloud-based. In order to generate an image from another image or text prompt you needed to connect to the service and generally there was a record of the work kept, so it could be tied to a user and a time. With Stable Diffusion all you need is a decent GPU and a bit of time. Ars Technica found an RTX 3060 12GB can generate 512×512 images in about 10 seconds, while a 3090 Ti can do it in four.
Stable Diffusion is open source and already people are making use of that to develop their own tweaks, to give the pictures a different flavour than the main branch. If you change the ~5 billion publicly available pictures it uses as a training set you can get very different results. It is also capable of looping back on itself, you can feed an image it generated back into Stable Diffusion to change or improve it. Hackaday started with a hand drawn picture of the Seattle skyline, which in a few generations turned into a realistic looking photo complete with an alien starship.
There are a lot of amazing things you can accomplish with AI image generation but there is a problem with it as well. Stable Diffusion’s license forbids its use for many nefarious purposes, but there is little detail in how that could possibly be enforced. We can hope that they’ve included some sort of metadata which would allow you to determine if the picture you are looking at is likely real, or AI generated.
It is impressive in it’s realism, that picture on the bottom right takes an outfit and body type which wouldn’t work outside of a cartoon, and rendered it into a very believable cosplayer. This could certainly cause many problems in the short term; who knows what this will mean in the long term.