Accelerating Character Design with AI-Powered Tools

About Artbreeder

Artbreeder is an incredible tool. It’s still the early days, but others and I have already started using it in our production concept art workflows. Artbreeder is a Generative Adversarial Network that allows artists to ‘sketch’ as fast as their imaginations. It draws from several massive datasets that have been carefully art directed and classified, using that information to generate novel imagery. You can upload your own picture of a character’s face, for example, and use its deep-learning networks to generate new features, shapes, colors, expressions, ethnicities, and much more.

As I already discussed in my previous article, when I’m designing a character for a film, I go through a lot of different ideas but I usually don't have enough time on the production schedule to explore everything or take risks in my design. But with Artbreeder AI, I can speed up the production and give myself more time to take those design risks and iterate on character's age, ethnicity, color, style, expressions, shapes, and more, all in real-time.

For 80 Level readers, here’s an Artbreeder special discount link if you want to try the tool out in your own workflow.

Discovering AI Technologies

I didn’t know that this interest would lead me to AI at the time, but early on at school, I realized the value of iteration. When I’m designing, I want to try out hundreds of different design options for a project. And if I have the luxury of time/budget to do this, I found that the further I go, the better the ideas I can come up with, the more iterations of the idea I can try. Over the years, I accumulated a variety of different tools that allow me to speed up the design process so that I can try out as many designs as my schedule will let me. Some notable workflows are using 3D, digital paint, photobashing, and kitbashing, all of which offer great ways to try out many design options in a limited amount of time. AI was a natural next step.

At Aliza Technologies, we started using GANs for animation, and immediately I knew that the ability to generate novel imagery based on art-directed datasets would apply to design as well, so in 2019 we started looking into building one for designers. Ultimately, we decided to go all-in on animation and so we dropped the design research, but at that time I had discovered Artbreeder and started a dialogue with them, and now have the pleasure of working with them and their software.

Can AI Replace Artists?

I definitely understand the fear of AI replacing us. But I think it's akin to worrying that the printing press would take away the jobs of scribes. Or that Photobashing would take away the jobs of painters.

AI is a pretty misunderstood term. We call it ‘intelligence’, but it's not quite the right term, as its ‘intelligence’ is just a series of programmed logical functions. Machine learning is very different from human learning, just as machine image generation is different from a human drawing. But when you combine these two, it elevates the potential for the human designer. Just like an illustrator using Photoshop to draw, there is no worry that having an ‘Undo’ function will cause the artist to be more productive and therefore more artists will lose jobs; on the contrary, it elevates the entire industry, causing us to be even more productive, lowering the barrier to entry, and giving more and more people the joy of creativity. Digital painting, 3D, all these have allowed smaller teams with less budget to create world-class content. And AI is going to continue this trend. Don’t worry, there is no spark of imagination behind it unless there is a human to guide it. And even if someday that happens… will humans even care about a solely machine-created medium? I think that’s doubtful. After all, no one wants to buy my photograph of the Mona Lisa.

Can AI Replicate Stylized Art?

The key to machine learning is massive amounts of data to learn a subject or a style. So for example, we have great GANs for human faces, because that is the most abundant type of imagery that exists, by an order of magnitude. There are billions of real pictures of human faces that have been learned by networks by now (see Nvidia’s ThisPersonDoesNotExist.com), and so the networks have gotten incredibly good at generating human faces, in a ‘real’ style. But there’s only one Picasso, only one Syd Mead, only one of you and one of me. And so, if we wanted to get a GAN to learn our style, it wouldn't be able to do it. None of us (okay, maybe Picasso) has generated enough imagery for a network to learn it, and then apply it to novel image generation. This is a massive limitation, but it's also a massive opportunity. I think an important commodity in the future is going to be style datasets hand-crafted and art directed by specific artists, using very advanced classification systems to generate the style. And even then, you’re going to need that spark of human imagination to guide the image generation to create the right image for the job.

New Technologies are Changing the Industry

I think we’re seeing how an entirely new medium is being born. This means that the barrier to entry for creating world-class content continues to lower, and more people get to share the beauty of their personal imagination with the world. 

As an example, there are many fashion designers who need to put more time into learning patterning than they do, say, into drawing human faces. But wouldn't their fashion illustrations benefit from great portraiture? Now they’re able to dual-track their skill-set and get just as great imagery out of both processes, without needing to double the time to learn. Obviously, I still highly recommend drawing for a foundational skill set, but my point is that the further we push this technology, this new medium, the more people are going to be able to create world-class content without needing a million-dollar budget.

It's an exciting time to be a creative.

Xander Smith, Concept Designer/ 3D Artist and Lead Artist at Aliza Technologies

Interview conducted by Arti Sergeev