Ya know, like:
I was hoping to make my dream come true with Advil & Legacy, alas, neither is available, so I had to request Colgate & The North Face (published and not that interesting, oh well).
Don’t read too much into “engine”, I think dude is probably just using Stable Diffusion or DALL•E and attempting to get something interesting. Most of them look like about what you’d get typing in basic brand terms and nouns. Still very clever and fun.
(via Rob Walker in FastCompany)
Ever wonder exactly how these AI image generators work? I think we all know they are trained on massive image sets ethics-be-damned-ed-ly scraped from the internet. But then what? What does training really mean and how does it craft new images from this pile of data?
The Washington Post did a great job of explaining. During training, noise is added to the images, obscuring them, and the job of the engine is to still recognize the images despite the noise, and also learn to de-noise it. They get super good at de-noising images, of which they also know what the original description of that image is. Then, they are provided with a totally randomized square of noise, given a prompt that a user like you or me provides, and told to de-noise it. Even though the original is just random nothing, the engine will try its best to de-noise it based on everything it knows about the words in the prompt, thus giving us these weird and wonderful creations.