We recently conducted an interview with HomeStratosphere discussing AI Artwork in the realm of home decor.
The original article can be found here: https://www.homestratosphere.com/ai-canvas-art-blog-interest/
What exactly is AI Art, and how does it differ from other computer generated artwork?
Over the last few years, there has been a real explosion in the field of artificial intelligence. A lot of it has been driven by increased computing power, but also various ‘small’ discoveries and enhancements which made AI perform a lot better. These neural networks, to grossly oversimplify, allow computers to figure out their own solutions to problems.
In the case of AI Art, we’re asking an AI system to create an image that resembles impressionist scenery, renaissance portraits, abstract art, etc, using only random numbers as an input. With enough training and examples, they figure out how to create that content.
A lot of computer-generated artwork of the past relied on computer artists and 3D modelers using software to manually place objects in scenes and render them. Or in some cases just smart programmers making algorithms that could draw very complex patterns. These approaches give a lot of control, but ultimately a human had to tell the machine precisely what to do, in order to get a result.
With AI, a human is still involved a bit, but ultimately the machine figures it out, without needing a human to tell it all the steps.
What are the pros and cons of original AI Art versus other options for home decorating?
I’d say the biggest benefit is originality and cost. Every piece is one of a kind, so the piece that you put in your home is quite literally the only copy that will ever exist. I think that’s such a cool feeling to have – your very own private collection of art.
And because it’s machine-generated the costs are very reasonable. We can provide a 3-foot x 3-foot wood-framed canvas typically below $400 (including delivery), whereas original paintings of that size are easily pricing at $2,000 to start and go up from there.
The biggest drawback is probably tied to texture – with an authentic oil painting for example having depth to the brush strokes or an abstract artist that uses yarn, plaster, or other unusual materials to create their work – we aren’t able to replicate that (yet). However, the canvas itself is textured, which looks quite nice, and we think an “original” print hands down beats a print that’s been made thousands of times – like the prints you might buy in a store or other mass-produced art sites.
What kind of art styles do you currently sell? What other styles are you thinking about?
We currently have 6 major collections, including Impressionist Landscapes, Impressionist Portraits, Retro Abstracts, Moderna Abstracts, a new abstract collection called Windstrokes, along with a photorealism gallery called ‘the Alleys of Old Europe’. We’re always looking for new content, so I think there will be new releases over time within those existing collections, but there are some new AI models that might be better suited to “surrealism” or even more wildly ‘unique’ runs.
How long have AI systems been around that can do this type of work?
The earliest AI system doing this type of work was created by Ian Goodfellow in 2014. The network he devised was called a generative adversarial network, in which one network learns to create images to fool another network designed to detect fakes – over time the generator gets very good at making new content because both networks battle against each other and must get continuously better.
My memory may be hazy, but at the time I believe it was able to generate black and white text that was 64 pixels by 64 pixels. Since then, a lot of advances have been made – partly from just brute force computing power, but also from the more intelligent network design.
Have there been any big ticket AI generated art sales at galleries?
Yes, very big-ticket. In late 2018, an AI-generated portrait titled Edmond de Belamy was sold for $432,500. It was kind of unique in that it was the ‘first’ AI artwork ever auctioned at a prestigious gallery like Christie’s. Who knows how the market develops over time – but there could be ways that AI art appreciates significantly, especially as works that are created by studying rare private content – artists trained on that material could be a lot more unique in their generation capabilities.
Technically speaking – how does it actually work?
Without getting too technical, what you do is create two AI networks. One network gets rewarded whenever it can detect an image as being “fake” out of a sample of images that are a mix of “fake” and “real” artwork. A second network is rewarded whenever it creates a fake artwork that successfully fools the first network.
When you make these networks play millions of games against each other, the generator network gets very good at creating artwork that has the colors, style, and features of the subject matter. We can then use that generator network to create new images.
When you say that the AI system learns – what does that look like in practice – can you give an example of how a poorly trained artist “matures” into a better artist?
Sure – so to be clear you don’t have an AI artist making brushstrokes on a screen – it is maybe better described as the machine ‘hallucinating’ an image all at once. So with no training at all, the machine is making colorful blurry noise. As it practices a few thousand tries becomes better-defined noise. Let it train a hundred thousand more times and you can start to see a coastline with a clear break between water and the sky.
As you start to get to a million attempts, the machine starts to create water and sky patterns, with some foliage and grass, and some reflections – although the reflections are inaccurate. Lastly, you can see below it decided to drop the orange-colored forests and focus on more highly defined living and dead trees on the coast. In addition, the water reflections and ripples become much more subtle and accurate, versus the sky which retains large pillowy clouds.
In theory, you can keep letting it train forever, and keep adding new materials to further bolster its capabilities.
Does AI art look pixelated / what resolution is it printed at?
All of our content is at least 300 DPI, which means that all the detail of a 1200 pixel image (shown below, some grass from Meadow One) is printed into a 4-inch x 4-inch piece of canvas. On some of the smaller size frame options, that gets printed into an even smaller space. So you won’t see any jagged pixels in the print at all.
Sometimes the AI artist may pick up certain styles which have different levels of detail. Just like a handcrafted piece of furniture, we kind of look at those as the natural variation of the artist and its style. So you can see below the rocks and grass from a piece called “Cliffs in the Distance” where the artist transitions between a smoother style on the right-hand side, to one with more granularity at the top and left.
It looks great, and remains extremely high-resolution, but is just a natural transition from a softer ‘stroke’ to a harder ‘stroke’ in the image.
How long does it take for art to arrive?
Our fastest time for an order to deliver was 4 business days to Western Canada. But we typically suggest budgeting for 5-10 business days – it takes time to produce the work, frame it, ship it, and on the ground logistics in big cities can always take a little longer. Our view is that art typically isn’t an impulse next-day purchase so that delivery time is pretty reasonable – but if faster shipping is desired, we can certainly coordinate it if you reach out in advance.
How do you come up with the names for your artwork?
It’s a hybrid approach – some of the art is easy enough to name based on the content, and we’ve used that for most of the impressionist landscapes collection. But for the portraits collection, we actually built another AI system that was trained to look at medieval and renaissance names – and come up with its own names for the art – so both the artwork and the name of the person featured are entirely AI-generated.
Similarly, for the Alleys of Old Europe collection, we fed an AI naming system a list of towns in Europe, and it was able to invent new town names that sounded “old school European” but that to the best of our knowledge don’t actually exist. I think over time we trend toward fully automated naming because it saves time, and builds on the AI-generated art theme.
Can you do custom requests for materials / color / finishing?
Absolutely – for art with portrait/landscape/square dimensions, we have a good selection of sizes available, but can certainly entertain alternate dimensions with the same aspect ratio if you reach out in advance. Alternate finishing options, like acrylics or metallic, could be arranged as well, although they add a bit to the cost.
Do you offer any discounts / promotions to pricing?
We have some seasonal sales from time to time, and readers of this article can get 15% off during July and August 2021 using promo code STRATOSPHERE.
How does AI Art change the relationship with human artists?
Some artists understandably might see the evolution of AI-generated artwork as threatening, given the power of these machines will continue to grow, and the technology advances. But we think there are very unique collaborative options available as well. For example, in the future, an artist might paint several works of art as a starting point, then have an AI system study their work, and generate brand new “originals” in the artist’s own hallmark style.
So the artist could produce a lot of completely original prints in the same style as their physical work. Similarly, we could imagine an artist using an AI system to brainstorm ‘sketched’ content, from which a human artist could then use as a starting point for a new original oil painting. So I think there is a lot of opportunities for humans to use these systems as tools to supplement their own work.
How do you gather source material? Can it learn any artwork?
AI systems can learn to produce anything as long as you have enough examples of material for it to learn on. There are lots of neat examples where GANs have been trained on thousands of human faces to create new, photorealistic people that don’t exist. We tend to gather a lot of content from the internet but are branching out into sourcing original new content for training which is harder to replicate and provides a unique and differentiated style.
What are the biggest limitations?
AI systems are very good at learning what they’ve seen – but aren’t so good yet at imagining outside-the-box content. So you won’t get an AI system that can paint impressionist style sportscars just because it can do impressionist style landscapes. These systems also take a lot of computing power and there is quite a bit of human effort in curating the content they learn from – so you want to be sure the content is of good quality and has enough variety to allow the system to generate interesting outcomes.