Several years ago, a neat article in Forbes appeared:
I found it only because of my perpetual desires to find a way to design fictional characters and personas for various story ideas. I’ve been searching since 2013 for the perfect one, but they all have some shortcoming.
The most technically robust character creators were part of dedicated games, but of course you actually needed those games to get the experience. More than that, since they were parts of games, the base character creators very often did not have all possible customization items from the start— you’d have to buy and unlock more items and accessories as you played the game. What’s more, since games are often thematic, you may not be able to create the exact kind of character you want if there are any specific details in mind that aren’t also available in the creation system. But all in all, retail games have the best graphics of them all and you can usually do a lot more with them.
Free online character creators came in two packages: graphical programs and dress-up games. The former, which includes stuff like Mixamo, certainly have more ways to pose but have very few customization options since you either needed to download extra packs from their store or had to design it yourself.
The latter has always been the easiest. Dress-up games are basically just flash games where you dress up an avatar, typically designed like a doll, superhero, or anime character.
The big problems with dress-up games is that they are very often thematic and the art is dodgy— considering they’re made for flash game sites and offer little to no financial restitution for their creators, art assets are typically ready-made and low quality. There’s usually only one perspective— full-frontal or, less often, quarter-turn. And save for the best ones, you can’t edit any aspect of your character’s body outside of token masculine or feminine features since these are indeed dress-up games. In other words, you get what you pay for. You decided to go the free route instead of commissioning an artist, so you can’t complain that your character looks cheap.
With the rise of GANs, this may change in very short order. Indeed, it is entirely possible that we are within a year or two of a true “character creator: the game”. What’s more, the capabilities of such will be far beyond anything we see today even in the highest quality character creation systems.
This is due to three important factors:
- Text-to-image synthesis. In this theoretical game, you won’t necessarily need to fiddle with sliders, attachable items, or presets. Instead, you could type in a description into a box and near-instantly get your design as an output. Say I want to design an anime-style girl with jet-black hair, blue skin, pink eyes, and wearing such-and-such clothes with a devil tail and steampunk wings. Normally, I’d have to go through a series of different menus starting with the basic body type, then the hair, then the face, and so on and so forth. Here, that simple description alone will generate an image. If it’s not the one I want, I can keep generating them until I find one that’s at least close enough and then go in to edit the finer details if needbe.
- Variable artstyle or graphics. If I want to create a character in the style of the Simpsons, I either need to commission an artist who draws in that style, find a flash game that allows me to edit a character and hope they have what I want, or learn to draw myself. And what if I want another character in 3D but in a janky, Sega Saturn or PS1-style polygonal graphics? With this theoretical game, this won’t be much of a problem either. As long as you give the GAN something from your preferred style, of course, it could conceivably give you a character that’s minimalist, blocky, cel-shaded, photorealisic, and everything in between. For example, if I want to generate a character that looks as if it were drawn by the mangaka Akira Toriyama, I could. If I wanted a character “drawn” by cartoonist Tex Avery, I could very well get one. If I wanted a photorealistic avatar, I could have that as well. This could be used to create that generated comic I talked about before, and it could also theoretically be used to create character models that modders can insert into old games.
- Unlimited flexibility. Because of the aforementioned aspects, there’s no limit to what you can create. You wouldn’t need to worry about whether or not certain assets are in the engine— as long as you can provide the GAN with some representation of that asset, it’ll be able to translate it onto your character. For example: almost all character creators don’t have wild, electrical, Super Saiyan-esque hair. But if you can give the GAN images like this, it will remember that design and even be able to fill in the blanks if it’s not exactly what you want. What if your character is supposed to have neon glow-in-the-dark hair like this? You’re welcome, I’ve just given you all you really need for your character in the future.
The possibilities are endless. And of course, if this GAN can create a character, it can obviously create other things. But this is what I’m focusing on at the moment.
I think we’ll see early variants of it this year, building off the recently released StyleGAN network. We’ll even see some text-to-image synthesis, which is well within the powers of current day AI.