The monstrous cats of Pix2pix, backed by TensorFlow, didn't take long to travel across the Web, and obviously it was a matter of time before someone took that code and decided to create a different solution. Artificial intelligences are very good at recognizing objects and faces, but when it comes to generating an image from a simple group of data, they still have a lot to learn. What would be a 'simple group' of data? Draw a face freehand in the box, and let Pix2pix sweat a little.
Training and Interpretation
No matter its level of complexity, the key to any artificial intelligence is training. If this sounds a bit like brute force, it's because in a way it is: If you want an artificial intelligence that is formidable at recognizing oranges and lemons, the most effective path is to feed it millions of images, let it make all the mistakes it needs, and watch how it improves. This process demands considerable processing power, and the latest developments aim to reduce training time, but at the same time there is the possibility of taking an algorithm out of its comfort zone and forcing it to make strange interpretations.
One of the best examples with free access for the user is Pix2pix. The first time we saw it up close was in February, when programmer Christopher Hesse created a port compatible with TensorFlow. Its goal was to interpret anything drawn as if it were a cat, and when I say 'anything', let's say even a Beholder could gain fur and cat eyes. This new version focuses on human faces, and basically depends on how complex the source is. If we look at the reference image, its result is very good, and even when we alter it a bit, it tries to apply what it learned, say, creating a neck where there is none.
Needless to say, Pix2pix's little sanity flies off the cliff when we force it to process simple doodles. If the priority is to see everything as a human face, very funny things can come out, or monstrosities worthy of master H.R. Giger. Unfortunately I don't have any advanced technical data on this generator beyond its creator, NPO (the public broadcasting company of the Netherlands), and the dataset, based on images from journalist Lara Rense. Try to draw something, and share the results.
