To quote Google CEO Sundar Pichai: AI is “more profound than fire or electricity.”
To back up this claim with incontrovertible evidence: here’s an AI tool made by Google researchers that turns doodles into weird monsters. What could be more profound?
It’s certainly fun, anyway. The tool is called Chimera Painter and uses machine learning to generate imagery based on users’ rough sketches. This sort of dynamic is becoming a relatively common one in machine learning. Nvidia has done it with landscapes before; MIT and IBM did it with buildings; and now Google is ... doing it with monsters.
The team behind Chimera Painter explained their methods and motivations in a blog post, saying the idea was to create a “paintbrush that acted less like a tool and more like an assistant.” Chimera Painter is just a prototype, but if software like this becomes common it could “reduce the amount of time necessary to create high-quality art,” claim the team.
The researchers gave themselves the challenge of creating artwork for a fictional fantasy card game, in which players combine features from different monsters and battle them like mutating Pokémon. They trained a machine learning model on a database of more than 10,000 sample monsters, which were themselves in part procedurally generated using 3D models rendered in Unreal Engine. Each image is paired with a “segmentation map” — an overlay that divides the monsters into anatomical parts like claws, snouts, legs, and so on.
Once the model has been trained on this data, users can then paint their own segmentation map which is then rendered using photorealistic textures. If you load up Chimera Painter you can see some of the preset monsters and they’re impressively cohesive. However, painting one yourself takes more time and effort than you might think. Our attempt below, for example, looks like a knock-off Gruffalo depicted using mud. It’s monstrous, but not necessarily a monster.