A new Tensorflow port of pix2pix is able to take your crude cat sketches and transform them into photo-realistic cat abominations — that is, photos that look like cats at first, but upon closer inspection are instead trippy piecemeal simulated versions of real cats. The system is also able to generate fairly realistic building facades, shoes, and bags using both preexisting sketches and ones you create yourself.
The demo is available for anyone to try here. The system works, in the case of the cat tool, bag, and shoes, by inputing a sketch, processing it using artificial intelligence technology, and then presenting the output image, which is a more or less photo realistic version of that sketch. You can create your own sketches using the line tool or modify existing sketches using the eraser tool.
The AI is able to generate these images thanks to training performed using thousands of existing images of cats, buildings, hand bags, and shoes. More than 2,000 cat stock images were used during training, for example, as well as 50,000 pictures of shoes via Zappos and about 137,000 photos of hand bags.
This isn’t the first time we’ve seen machine learning demos able to create trippy and/or creepy photos using sketches and commands. There’s Google’s dreamy AI landscapes, for example, and there’s even the horror-image-generating “Nightmare Machine” AI from MIT that rehashes ordinary photos as scary renditions.