![](https://crypto4nerd.com/wp-content/uploads/2023/02/1mzgYE6ZxiQ-6RuyyYIUZSA.png)
I’ve been looking at AI for creative purposes for a while and wanted to jot down some of my findings.
Why as a designer, would I want to look at AI? On the surface artists and designers have a lot to be scared of when it comes to this kind of thing. To have images generated out of thin air in minutes from commands written in plain English into a text editor evokes the same feelings the pre-industrial revolution textile workers must have had when they saw machines replacing their jobs. Will we need photographers, illustrators, concept artists, photo retouchers and animators any more? I feel I’m going to have to make a choice pretty soon: become a Luddite or embracer of technology.
My own reasons for looking are a bit diffrent to others. I’ve had a problem with my vision that’s making creating design and artwork difficult to do. I was looking for something that would still use my creativity and visual sense but that avoided the necessity to sit for hours staring at a screen to perfect an image or layout. Something to take to load off, rather than do my job and yet still produce work that was under my control. After all, that’s what happened in the late ’80s with desktop publishing. Designers and typesetters didn’t become obsolete — the ones who embraced the new technology benefitted from it enormously. As well as being a designer, I was training other creatives in the early ’90s, helping them to move over to Mac workflows and I saw the “lightbulb moment” in every one of them as they suddenly realised that their ability to create had skyrocketed.
However, even before my eye problems, I was looking at AI for quick ways of getting specific images without trawling through stock sites or setting up my own photoshoots. I know enough photographers to feel uncomfortable doing this, so there was always the intent to get a photographer in to do the “real thing” if the client liked the concept. I find writing blog posts pretty easy, but constantly coming up with new ideas for them is draining and AI has given me the inspiration I needed to get started a few times as well.
My reason for not going 100% with AI produced work, be it visual, musical or written, is that “you can always tell the difference”. But after looking at the thousands of examples of AI produced art that have appeared during the beta tests of various systems I began to wonder whether it was only me or a photographer or an artist who could “always tell the difference”. Lots of other people seemed happy with what was coming out of the machine. In fact some artists were being accused of using AI because some of their original work came out looking like AI imagery that had been generated by other people, asking for it to be rendered in that artist’s style.
I wondered two things:
- Is there an equivalent of the Turing Test for AI’s creative output?
- Could I have an AI system that was an analogue of myself and knew me so well that it produced exactly what I would do, but without me actually having to do it?
I decided to use Midjourney and devise some tests to see if it knew its stuff when it comes to the subtleties of art and design and also to see if it could learn my own style. As all machine learning depends heavily on interpreting past data and knowledge to create new work (pretty much like people, really) I’d expect it to be better at imitating well-known styles rather than my own as there would be more references so I was also looking at how it coped with more vague prompts as opposed to more precise ones.
Prompt: “five table lamps in a row in different styles: 1. Art Deco, 2. Bauhaus, 3. Art Nouveau, 4. Memphis, 5. De Stijl”
This is something you might ask schoolkids to do for a design and technology project. It involves research, and then a creative interpretation of what you have learnt.
I asked for five table lamps “in a row” which I didn’t get on any of the options and in number one, I only got four lamps! However, the layout it picked and the 3D rendering was really quite impressive. Some looked like table lamps and were actually casting light. I have no idea where all of this came from.
It’s not really clear what materials the lamps are made from. This is important because it looks as some of the styles use materials that weren’t in existence when they would have been designed — for instance, brightly coloured, glossy acrylic would have really been duller bakelite on the Nouveau and Deco objects. A design historian or collector would pick these details up.
As for the styles, I think that we can safely say that there’s a bias towards De Stijl. It’s not because it’s the most recent style on the list (that would be Memphis), but I think if you were describing it to a child, De Stijl is probably the easiest to get across in words (bold, primary colours, black outlines, simple shapes). Maybe this is why it’s more of an influence?
I chose these five styles as there’s a line of evolution, with connections between them all, a well as obvious and more subtle differences. The middle-back one on the second image is interesting as it has influences from all of the named styles.
Across the board Art Nouveau (the oldest and also most organic and natural style) seems to be the most neglected overall, which is a shame as it was a chance for the machine to show us it’s not just about engineering and straight lines and that it has empathy with softer, natural forms.
Bauhaus was always going to be difficult because it’s more of an approach to design rather than a specific style but front-left in both number two and number three would probably be enough to make an antique dealer take a second look.
Interestingly, we have some branding going on in number two — it’s gone ahead and designed some kind of a logo, which I never asked for!
Prompt: “The Scream but happy.”
There are five versions of Edvard Munch’s The Scream, so I left the prompt open to see which the system would pick, if any. I’d immediately think of this painting given this prompt, but it wasn’t guaranteed. Being such a well-known work, it successfully predicted my thoughts. It seems to have preferred the brighter tempera versions to Munch’s litho print, pastel and oil ones, which might be the “happy” influence coming in.
However, it failed big time in the happy department! The colours and style could be described as more lighthearted, but I was expecting the content to have been looked at with the actual scream possibly being turned into a smile or laugh (at least, this is what I would have done). What it did was put different faces into the picture, all of which had a wide open mouth.
The misery present in the original (all five, in fact) is so obvious to a human viewer but seems to not have any relevance to the machine. I think the word “Scream” was given too much emphasis in finding a useable facial expression rather than just being the title of the original image. “Happy” went towards the style of the picture, rather than the emotion in it.
If you read up on the painting, the background of The Scream is as important as (if not more than) the figure. This makes number three wrong in almost every way.
A variation on the fourth picture gave this, which was just…weird.
Prompt: “Portrait of Andy Oakey in Bauhaus style in charcoal.”
Slightly vain, this one, but I wanted to see what happens if you name somebody in the search without giving a visual reference. I also wanted to see what it made of “Bauhaus- style”, as a Bauhaus style doesn’t really exist but you can look at something and say “that’s Bauhaus” without knowing why.
The images were interesting. The first images of an Andy Oakey on Google aren’t of me (ego crushed), but of someone wearing glasses. Lo and behold, two of the resulting portraits are wearing glasses and resemble my namesake. However, two of them are baldies like me and resemble my LinkedIn profile pic. Strangely, number three resembles a photo of me that’s on Facebook when I had hair. Strange because in the original of that I was wearing glasses myself, which have been removed!
The Bauhaus style seems to have manifested itself in the clothing rather than the artistic influence, although number four has a really nice background. All have a bit of a noir feel, which has been helped by the request for charcoal in the prompt forcing them to be monochrome. I do like number four, but he’s way too handsome to replace my profile pic. with. The cheekbones have been emphasised in all of them but show up most on this one. Requesting variants of one and four came up with an almost identical set of images, which was disappointing.
However, a variant of number three came up with this lot. Not particularly impressive at first glance. One looks like a bare-knuckle fighter out of Peaky Blinders and the next two are just a bit, well, nothing, really. However it struck me that number four really resembles my grandad from an old wartime photo. On this one, there’s also a words at the bottom. This has me baffled. Am I missing something? Does this resemble a particular artist’s work so much that it’s got their signature? If Danyy Dary is around anywhere, please get in touch! It has an almost Art Deco, slabby, Tamara de Lempicka feel about it and I also love the sunlight and shadow window frame behind.
Prompt: “[my Linkedin profile pic URL] In a splattery graphic style.”
Here I wanted to see if I could have more influence over the original image by giving it a link to a specific photo as a start point (I chose my LinkedIn profile pic.)
My second goal was more interesting. I wanted to see if I could get a similar style of illustration to my own and to see what my own hands came up with vs. that of the technology. If I can train an AI system myself to do work that looks like it’s mine, is it still mine? Huge questions, and a whole new article in the making.
So, first answer was that yes, I can point it to my own photograph. However, Midjourney likes to mess with things, so my first attempt either took my photo and messed with the angle, or took a very similar picture of someone that was face-on. It also appeared that it had reflected the face to make it symmetrical and then added shadows to one side. No face is symmetrical. In fact that’s what makes people look human. It also changed my shirt! This made me think more that it was merging different images together. Also, is that paint or a terrible combover on number one!
I requested some variations on number two. The symmetry and the weirdness of the eyes made me not too keen on these. I liked the Jackson Pollock-ness of the thrown paint effects and how they interact with the head, but I think number four looks like I have a terrible head injury. Most AI’s rules usually block any requests for actual gory content, but maybe this got through the filters by being the wrong shade of red?
Keeping the same base image reference, I changed the prompt slightly to add some more details (“graphic pen, line drawing”) to more match my style. Well, that got interesting. It was much more faithful to the LinkedIn profile photo despite reversing the image (presumably for lighting purposes) and my drawing style. However, I’m now suddenly much younger, have interesting haircut and, well, I’m a girl!
The subject of gender hadn’t even occurred to me, to be honest. I’d just assumed that the picture of me was male, because I knew it was. What does AI see as masculine and feminine when you only have the subtleties of the human face to work with? This could be a minefield.
I decided to modify the prompt again this and added “masculine” to the prompt to see what it did. I also added “bald” to cut out the combover at the same time.
It seems that masculine to Midjourney means a more bulbous nose, thinner eyes and a bigger chin. I also seem to favour looking to my left which is opposite to the previous set. Again, this could be to do with lighting needs, or it might be going back to the original source image that was facing that way. Or it realised that mens’ shirts button up on the opposite side?
The shirt is back but the likeness isn’t great, with number two looking like a shifty Tom Cruise; maybe that’s what Midjourney thinks of as masculine? On balance three is best and so I went for some more variants of that.
I’m now concentrating on the illustration style to get something that looks like it was by me as well as of me.
Two is the most interesting, style-wise and number three’s grey background with the cut-off shoulder also looks quite natural. In both of these, the shirt is more successful than the face in terms of looking like an illustration.
More left/right jiggery-pokery in the next set, but the illustration style is getting better.
In this last crop, I think two and three are the most natural-looking illustrations that could have been drawn by hand (mine, specifically) and also capture a near-ish likeness.
Finally, here’s the ‘real thing’. My own pen-and-ink attempt at answering the brief:
Would you have been able to tell the difference between this and the machine’s efforts?
Answers to my original questions:
- Is there an equivalent of the Turing Test for AI’s creative output?
The Turing Test puts a human and computer up against a third interrogator (another human) who asks both of them questions to see if the interrogator can tell which one is real and which is the machine. The answers are in text format (speech synthesis wasn’t great in 1950, so the Dalek-like voice would have been a giveaway). For AI art, music or copywriting I think you’d need an expert in each particular discipline to check the results properly. The example of the lamps being made from the wrong materials highlights this — a general consumer might not spot something like that. - Could I have an AI system that was an analogue of myself and knew me so well that it produced exactly what I would do, but without me actually having to do it?
I think that with some time, it would be possible to have a standard set of words in your query to ‘set the scene’ and map out exactly how you work, what techniques you use, etc. If this baseline was always in every request you made it would keep things consistent. However, creativity is often about change and doing surprising things rather than keeping safe. David Bowie is a perfect example. Every Bowie song is recognisable, but there are so many styles and variations in his work. Asking AI to come up with a Bowie song will never be able to create the next Bowie song — just a slightly disturbing but recognisable mix of his most well-known work.
So will AI replace AO? In some ways, I hope so. Freeing me of the chore of actually doing, whilst leaving me to just think seems ideal. It’s no different to being able to select a font from a menu as opposed to having to draw all of the letters yourself. I’ll never stop enjoying the act of painting and drawing or taking photos, but doing these to an exact brief laid down by a client, against a deadline and with a dodgy eye taints the enjoyment. The final self-portrait took me about 90 minutes to draw and paint. It took Midjourney 30 seconds. As a relaxing exercise it was an hour-and-a-half well spent, but in a commercial environment, that sort of difference in productivity could equate to thousands of pounds being saved over a day — I could do two, maybe three of these in a day, whereas AI could do that many in a couple of minutes if programmed properly (and also keep learning how to get better).
As with all technology AI can be misused, overused or embraced but it in the same way as the internet, or mobile phones eased their way into all of our lives, it can’t be ignored.