Meet Vincent, the AI artist bringing your sketches to life
What could you create if you had the skill, talent and vision of history’s greatest artists? Would you draw the surreal scenes of Dali imbued with the illusionary shapes of Escher? Perhaps a Warhol-esque screen print is more your style, finished off with flecks of Picasso’s cubist period to round it off.
Maybe you’d just make a mess. But what if an AI was bestowed with that power? What if you could draw upon that knowledge to improve your own artistic skills? That’s where Vincent comes in.
Plied with images of artworks from over 8,000 artists, Vincent is an AI developed by Cambridge Consultants to help create artworks. Working off the edges a user sketches onto its interface, Vincent dynamically paints a picture.
“If all the artists we trained [Vincent] on were forced to get together and collaborate on an image, this is a possible image of what they might draw,” explains Monty Barlow, Cambridge Consultants’ director of machine learning. “It’s not perfect, but the more we draw, the more of a sense it gets [of what we want to draw].”
Draw a straight line on an empty canvas and Vincent imagines you’re creating a horizon. Plop in zig-zag peaks and it perceives mountains, downward lines could be rivers, ice sheets, or shadowing. Blobs in the sky become clouds and items drawn towards the bottom of the canvas are deemed to be in the foreground. Doodle below a ‘horizon’ and Vincent will shift the image’s focus accordingly.
“You’re giving it 1% of the input and it’s making the rest up. It’s somehow joining the dots. That, to date, has been very difficult to do.”
There’s more to Vincent than simply trying to help bad drawers create wonderful works of art, it’s about understanding how humans and AI interact with one another. If given a small bit of information, could an AI finish off that thought or creation for you?
“[Vincent] is still giving you control, even though it may feel like you don’t have much,” says Barlow. “This is your picture, you’re defining what it constitutes of and, the majority of the time, if you do the same image you’ll get the same picture back out.”
For the most part, that’s true. I draw a couple of very similar shapes and a similar image is produced both times. Vincent can tap into numerous neural networks to switch up its art style. You can think of these networks as different creative brains, each one trained on the same set of data, but with its parameters tweaked to prefer certain aesthetics. “Some of [the neural networks] are a bit random,” Barlow explains, “one of them just wants to draw body parts, while another does better with simple bold shapes.”
Vincent is remarkably capable. It makes intelligent decisions on how certain things should always look. “We took edges off an actual piece of artwork and it’s remarkable how many of them automatically had pink lips,” Barlow explained. “In one image of a face in a computer screen, no matter how much you change it, it always sees a pair of big red lips.”
As impressive as it is, I have to agree with Barlow that the computer’s face is “worryingly consistent and absolutely horrid”. But it’s interesting because, despite the face being contained within the confines of a drawn computer screen, Vincent recognised it as a human face.
Vincent can also interpret more advanced sketches, such as that of a three-dimensional head. Here, instead of seeing it as a flat object, Vincent sees a sculpture of a face complete with heavy brow and deep-set eyes. “A child may have not even recognised [those curved lines] as a face, but there was enough skill in the shape that the brow was there and [Vincent] could see a chin and shoulders too.”
What’s next for Vincent?
AI collaboration is just one part of Cambridge Consultants’ research project, Barlow and his team are also looking at how you can use AI to squeeze a lot of information out of a small dataset. By using its 8,000 pieces of unlabelled artwork, Vincent has had to learn what makes each painting unique and then relate those elements to the shapes and figures you draw. It’s learning from the information it knows, rather than being explicitly told what each piece is and finding a solution that fits.
The research that’s going on as part of Vincent isn’t just limited to art or imagery. Its uses could extend far beyond into more practical circumstances that businesses could even adopt in the future. “Imagine giving [an AI] 2% of a wedding or party plan, and letting it tap into all the parties created by the 100-best party planners in history. It would spit out ideas near instantly, each one releasing all the parameters you requested.”
Barlow compares the process to that of the work Nvidia is doing with its AI testing for autonomous vehicles. “It’s a complicated area, you either have to synthetically make pedestrians jump out in front of vehicles, or take what footage you do have and do clever stuff like change night into day or people into billboards.
“The way we can manipulate and use data has really changed, even in just the last year or so.”
Only last year, generative real-time drawing could only be done on a 32-pixel square space. Now though, as Barlow explains, they can work up to sizes of 1-megapixel – in a year or two Vincent, or similar system could potentially be capable of photorealistic generative art creation.
Barlow’s research team have also just worked on a demo using a similar principle but trained on classical music. By feeding thousands of classical piano pieces into a neural network – again, unlabelled – it can determine if an entirely original piece of piano playing is Baroque or another type of Classical style.
If you’re wondering where Cambridge Consultant’s are going with Vincent, even they don’t really know. “Vincent just goes on exploring and touring schools,” reason’s Barlow, “but the pieces behind this are always evolving.
“It’s a journey of manipulating data, often not visual.”
Vincent isn’t about coming out with a defined end goal. It’s a project centred on learning and exploring the potentials of AI as a tool for creation – be that visual, acoustic or practical purposes like event planning or architectural design. A number of other projects have experimented with image-creation neural networks trained on labelled datasets, but Vincent is different. Instead of having labelled data fed to it, Vincent simply has a repository of images it knows nothing about. Using these, it recognises shapes and patterns to decide the rules of what should fill a shape or line drawn by a creator. It’s learning to understand what someone could be drawing or want to draw, rather than simply providing predetermined images that fit a shape in the best way possible.
You may be familiar with Google’s DeepDream project, have seen the “edges2cats” experiment that brings nightmarish cat drawings to life, or the AI that builds fake cities from sketches. These processes work on a similar basis to that of Vincent, in as much as they’ve been fed a database and can create from it, but the fundamental difference is that none of these projects are learning in the same way. You input what you want and it ten thinks about it and creates it. Vincent changes dynamically as you draw, it’s working away continually; working with you to create an image, instead of simply obeying your command. It’s an AI partner, co-creator, collaborator, not simply another input/output device.
Vincent is more than just an image creation tool, but it’s framed in such a way to make all the science going on behind the scenes more digestible and easier to parse. However, what is happening behind the scenes is pushing the boundaries of our understanding of AI as a tool.
“That future where you collaborate with an AI is coming, and there hasn’t been anything like this yet.”