Google's Chimera Painter Can Create Fantastical Creatures Automatically

Google has created a new machine learning model that can automatically create renderings from a user supplied creature outline. The trained machine learning model is called Chimera Painter. It's a machine learning model that can act as a paintbrush for a digital artist to reduce the amount of time required for creating high-quality art without impacting artistic choice.

Chimera Painter is a demo application that can add features and textures to a creature outline segmented with body parts labeled with labels like "wings" or "claws." When the user presses the transform button, the algorithm would create the artwork for them. Google believes the tool could be used to create video game-ready assets.

The images above and below were created by Chimera Painter and are for a digital card game prototype designed around the concept of combining creatures to create new hybrids that can battle each other. For the game, the player would begin with cards featuring real-world animals and make them more powerful by combining them. Google says the Chimera Painter algorithm was designed to create a large volume of artistic assets that can be combined naturally while retaining identifiable visual characteristics of the original characters.

The project team experimented with generative adversarial networks (GANs) that were informed using artist feedback to create images appropriate for the fantasy card game prototype. GANs pair two convolutional neural networks against each other with a generator network to create new images and a discriminator network to determine if the images are samples from the training data set. The training data set in this instance was a set of artist-created images.

Google's variant is a conditional GAN where the generator takes separate input and guides the image generation process. One thing to note about Google's project is that its approach was a departure from other GAN efforts that typically focus on photorealism. The images here are results from the Google Chimera Painter after the model was tasked with generating multi-species chimeras. Google integrated the best model into the Chimera Painter system, which created over 10,000 images for the set requiring approximately 20 minutes per image, saving vast amounts of time compared to manual image creation. Images were created using the Unreal Engine.