Talking about Google, among others, is talking about artificial intelligence and machine learning. Thanks to this we have a complete voice assistant, an operating system that learns from our use and an image recognition integrated within Google Lens. This is all very good but … What if you could turn any drawing into a fantastic creature?
The experiment that no one has asked for but is here to stay is called Chimera Painter and it is a machine learning model that allows turn our drawings into somewhat gloomy creatures. Through a antagonistic generative network (GAN), Google is able to make a doodle “come to life”. The tool also works from the phone, so we are going to explain what it consists of.
A generative network that creates chimeras
Chimera Painter is a machine learning model mainly focused on video game developers. However, is open to any user, so we can use it both for fun, and to understand what Google achieves through machine learning. This model creates images with a look very video game, in the style of ‘Magic’ cards or any other RPG with dungeon creatures.
Drawing on mobile is not easy, but my colleague Amparo Babiloni has created these wonders. Through a complete editor we can draw and select the parts of the body: head, legs, tail, nose, teeth, etc. From a simple drawing you will create a fantastic monster. Machine learning has its limits, and that is that either we work well on the drawing or we can recreate real dungeon spawns (although this is a large part of the fun). Beyond the fun and creative side, it is interesting to understand how this works.
Chimera Painter is a prototype of artificial intelligence to produce images ready to use in a video game. To give you an idea, it is a tool that has been based on the creation of a card game from scratch (a la Magic or Hearthstone). From this aesthetic of imaginary animals and beasts, created a creative environment on a visual level, later transferred to the network. GANs pair two convolutional neural networks with each other: a network that creates the images and another network that determines whether the images are samples from the training data set or not.
Through the images generated by Google with the card game, they have managed to train the neural network to transform any drawing into a 3D model with this aesthetic
To train this GAN, a set of color images was created, with creature outlines adapted from 3D models. Through this network they are able to identify and convert the lines we have drawn into a 3D model, although the results, as we have indicated, will depend a lot on our ability to draw.
If we want to test this tool, we only have to access Chimera Painter. We have several brushes, tool to erase and different colors for each part of the body. We already anticipate that it is quite difficult to draw from your mobile, but if you have pencil compatibility or a fairly generous screen, you may achieve surprising results.