Google develops 'Chimera Painter' that can generate realistic fictitious monsters with simple coloring
Google has developed ' Chimera Painter ' that can generate fictitious monsters with realistic textures by simply drawing monsters with lines of different colors for each part of the body.
Google AI Blog: Using GANs to Create Fantastical Creatures
Creating fictitious monsters that appear in games, etc. requires a high degree of artistic creativity and skills to express images. Artists with these skills create creative monsters while being chased by deadlines, creating various monsters that appear in the game world, so that gamers can enjoy fierce battles in the fictional world. Become.
' Chimera Painter ' was devised and developed by Google as a tool to help such artists 'work to create monsters'. Chimera Painter is a machine learning model generated using hostile generation networks (GANs) trained using hundreds of thousands of monster images created with 3D models.
Chimera Painter uses GANs that pair two convolutional neural networks with each other. One of the convolutional neural networks is the 'generator network that generates new images' and the other is the 'discriminator network that determines if the generated images are samples from the training dataset'.
A particular problem in creating creature generation tools using GANs is how to accurately output low-contrast parts of the image. The following image is an example where the part with low contrast cannot be output accurately. The part that is not output correctly is also very important for human perception, so a method to correct this was sought.
To solve this problem, Google has taken an artist-led, semi-automatic approach. This approach first asks the artist to create a monster using a 3D model. At this time, it seems that monsters were generated separately for each type, such as lions and hyenas. Next, textures are superimposed on the created model using Unreal Engine, and one is output as a realistic full-color image (left) and the other is output as a segmentation map (right) with different colors for each body part.
By this method, we succeeded in preparing hundreds of thousands of sets of 'full-color monster images generated from 3DCG models' and 'segmentation maps'. By training a machine learning model using a dataset that includes a full-color image and a segmentation map, the model will be able to correctly recognize and output parts of the monster's body.
In fact, Chimera Painter can be used by anyone from the following page. It is possible to automatically generate a monster image with a realistic texture just by drawing the monster's head (red), mouth (blue), wings (pink), tail (ivory), etc. in different colors. However, please note that the operation is slightly unstable at the time of writing the article.
If you click on the image below, you can see that just by adding a simple line, the appearance of the generated monster will change significantly.
The following are fictitious monsters actually output by Chimera Painter. It looks like it wouldn't be strange if it appeared in card games for smartphones.
Related Posts:
in Software, Web Service, Game, Design, Posted by logu_ii