Researchers at Google have developed a new AI tool called Chimera Painter that turns doodles into unusual creatures. This tool uses machine learning to create representation based on the user’s rough sketches. Before this, Nvidia has used a similar concept with landscapes, and MIT and IBM have produced a similar idea with buildings.
A high level of technical knowledge and artistic creativity is required to create art for digital video games. Game artists need to promptly iterate on ideas and develop many assets to meet tight deadlines.
The research team has aimed at creating a paintbrush that operates more like an assistant than a tool. Although the Chimera Painter is just a prototype, the team states that this software can lessen the time required to create high-quality art.
Prototyping the New Model
The researchers tried to create illustrations for a fictional fantasy card game, in which players link features from various creatures and battle them similar to mutating Pokémon.
Since the aim is to produce high-quality creature card images guided by artist input, the researchers have employed generative adversarial networks (GANs) to create creature images suitable for the fantasy card game prototype. GANs pair two convolutional neural networks. One is a generator network to create new images. The second is a discriminator network to determine if the pictures are samples from the training dataset or not. They have used a conditional GAN, where the generator takes a separate input to guide the image generation process.
For training the GANs, they created a dataset of full-color images with single-species creature outlines accustomed to 3D creature models. The creature outlines provided a segmentation map that identified individual body parts. Once the model was trained, it was tasked with generating multi-species chimeras based on artists’ designs. The best performing model was incorporated into the Chimera Painter.
Training to Generate Creatures with Structure
Using GANs for creating creatures could cause loss of anatomical and spatial coherence when detailed or low-contrast parts of images are rendered, despite high perceptual importance to humans.
Generating chimeras needed a new non-photographic fantasy-styled dataset with different characteristics like dramatic perspective, composition, and lighting. The existing repositories of illustrations may be subject to licensing restrictions, have conflicting features, or lack the diversity needed for this task. Therefore it was not appropriate to use as datasets for training an ML model.
The team has produced a new artist-led, semi-automated method to create a dataset for training ML models from 3D creature models to resolve this issue. It allowed them to work at scale and rapidly iterate as demanded. In this method, artists would create or obtain 3D creature models, one for each creature type required. They trained the ML model on a database of more than 10,000 sample monsters. These sample models were created in part procedurally using 3D models rendered in Unreal Engine.
Each picture is paired with a segmentation map. A segmentation map is an overlay dividing the monsters into anatomical parts like claws, muzzles, limbs, etc. A set of programmed scripts then take this 3D scene and interpolate in different poses, viewpoints, and zoom levels for every 3D creature model. This creates full-color images and segmentation maps, making the training dataset for the GAN.
The trained GAN is available in the Chimera Painter demo. Artists can now work iteratively with the model rather than drawing numerous similar creatures from scratch. Users can paint their segmentation map. Afterward, photorealistic textures are used to render the user-created map.