When I first saw the images Andy Wallace has generated with his neural net trained on Magic cards, my immediate thought was "wait, what card is that?" It took me a moment to realize that these aren't cards at all, but instead they are conglomerations or ideas of Magic cards. In the same way that the semi-famous Deep Dream algorithm looks at a lot of pictures of dogs to then turn other things into, well, almost-pictures of dogs, Wallace's Magic bot has taken visual themes, concepts, and patterns from Magic: The Gathering cards and algorithm'd up new images.
The origin of Wallace's neural network Magic art comes from a practical place. In an interview over email, he told me that the project started as a way of creating art for cards that are created and curated by the Twitter account @roborosewater, which features cards generated from an algorithm trained on the rules and text of Magic cards. "The text was generated via neural net," he told me, "so it seemed fitting for the art to be as well."
The most impressive thing about the images, to my mind, is that they all manage to look distinctly like they're from the same game. Even more impressive is that most Magic cards don't look alike in any way. Wallace feels a similar way, and he explained the kind of wonder of what the bot does in a beautiful way when he wrote this to me:
"Despite the seemingly disjointed input, the output from my bot still comes off as distinctly Magic: The Gathering, albeit a bit abstract. Rebecca Guay was always one of my favorite artists in the game, and I think some of her flowing style reads in a lot of these images. The pictures created by my bot lack distinct characters which are present in nearly all MTG art (even spells), but I think the colors and shapes still evoke a scale and mystery that feels right."
Magic: The Gathering is a game that often depends on feel: Does this spell feel red? Does this art feel like it fits this card? Does countering someone's spell feel bad? Wallace's art bot captures this core pillar of the Magic game. It understands, and reflects, the feeling of a card. And that's wonderful.
You can find more images from Andy Wallace's Magic neural net at his Tumblr. All of the images are originally sized at 126x96 pixels and are then enlarged in an image editing program.