The great AI illustration experiment

I've recently seen a number of posts on my feed showing off AI generated artwork. I really knew nothing about it, though it all looked amazing, so I decided to have a nosey and see what I could do myself. I can confirm it is really pretty impressive to be honest and some of the results are genuinely stunning - easily good enough to fool feeble minded art directors like myself that I can now create cool art. In case this has so far passed you by the way it works is you find an AI art generator (I used Midjourney but there are others), then you simply type in a command and the keywords you want included in your image and hit return. The AI generates a choice of four artworks for you in about a minute, of which you can choose one or more to upscale to full size artwork. For the purposes of illustration below I am showing you all four original artworks offered up by the AI before you choose one to upscale. Very simplistically speaking (I know no other way) as I understand it the AI achieves this by searching the web for public domain 'datasets' (I'm not sure if this just means existing imagery) related to the keywords and then it processes and blends everything it finds into one image. You would be forgiven for thinking this just comes out as some poorly chopped up collage, but it really doesn't. The results can be very beautiful, but before all the great illustrators I know and respect break down into a blubbering mess and go start an only-fans channel there do appear to be some major limitations that at least for now mean your jobs are safe. I know this because I ran an experiment...

The concept experiment

I wanted to see if the AI could accurately deliver a conceptual illustration where the exact context was fundamental to the message of the illustration making sense rather than just looking pretty. This is how I went about it. I art direct Heart Matters magazine for the British Heart Foundation and we always commission a couple of illustrations per issue so I chose some previously commissioned illustrations and typed in a clean description of the concept these brilliant illustrators created for me and then let I loose the AI. For this experiment I chose to ignore BHF brand guidelines on colour and clean lines as the experiment was about concept delivery, not style. I did this for five illustrations. I am well aware this is a ridiculously small sample group if I were writing a white paper, but I'm not doing that, I am doing an art experiment here, not testing a new drug where lives hang in the balance, so gimme a break. All five results are pretty darn clear: the AI creates interesting, even challenging and beautiful artworks, but it cannot follow specific instructions when you have a clear vision of what you want in your concept artwork. It is simply not capable of following instructions that are specific. I will unquestionably continue to work with and commission illustrations when I need them because the power of their creativity and their ability come up with visual conceptual solutions are way way beyond the capabilities of AI. All you amazing conceptual illustrators are safe - at least for now. Below you can see the five examples that formed my experiment. A crucially important thing to remember here is that I specifically keyed in very clear instructions, I didn't ask the AI to come up with the concepts. The crucial idea which makes the illustration work and convey the message in a thought provoking way are exclusively the result of the illustrators' creativity. The way these AIs work will not give you a concept or an idea, it will only give you a visual interpretation of the specific terms you input.

So what is it good for?

Glad you asked. What this is just brilliant at is things like fantasy or SF artwork where you can loosely suggest subject matter but don't need a specific composition. Use keywords like 'fire' 'angel' 'destruction' 'doom' 'scorched earth' and you will get yourself the most spectacular heavy metal cover art you have ever seen. Use keywords like 'elf' 'fantasy' 'forest' 'battle' sword' 'dragon' 'mist' and you will have the perfect fantasy novel artwork. Use words like 'planets' 'solar' 'moon' 'galaxies' 'spaceship' 'nebula' and you have the perfect SF book cover. One other thing these AIs seem to be very good at is high res hyper real anthropomorphised cutesy animals in a sort of gamer-y style. There are loads of them doing the rounds. It's actually a lot of fun and you do genuinely feel you have come out with a great piece of artwork and all you need is five spare minutes and the ability to two-finger-type a bunch of keywords. From other efforts I saw on the website you don't even need to be able to spell correctly. I can see some small record labels and book publishers really leaning into this so the future of these fields of illustration is looking shaky. Or is it?

Assets, it's all about the assets

Maybe the future is not so bleak after all, even in these areas. If I read the information on Midjourney's website (it's all still in beta by the way so potentially subject to change) correctly the AI is only able to do its job by collating existing web based public domain visual assets. It's unclear if it is able to distinguish copyrighted artwork, or whether it cares. It's also unclear to me to what extent the AI simply copies or mutates these assets which opens up a whole can of worms in terms copyright which I expect will make a lot of potential commercial users rather twitchy. You can use your Midjourney artwork commercially on a Creative Commons License, but the copyright question is not an area which can be ignored and they themselves acknowledge this in the Midjourney T&Cs where there is a specific section on DMCA and takedown claims. These concerns may mean that commercial usage of these artworks is going to be limited or at least questionable. If you're still interested different levels of membership to the site grant different levels of access to the AI and different rights too. It's worth noting that a high percentage of clients require specific rights levels including exclusivity - even just for a period; I may have missed it but this doesn't seem to be possible under the current rights structure offered by Midjourney.

Geek out artboy

If we ignore the commercial side of things it all becomes great news though. If you really like the sound of that aforementioned metal angel of fire bestriding a scorched earth battlefield and you wish you could find a poster showing that very image to put on your bedroom wall then there is nothing stopping you simply using one of these AIs to create the image you want, in the style you want (yes, you can do that, you can add anything like 'in the style of Boris Vallejo' or 'cinematic' or 'oil painting' to your image tags), downloading the max resolution version of your image (on Midjourney that is currently 1664x1664px but no doubt that will increase) and simply lobbing it over the hedge at your local print shop. I also fully expect that vast numbers of etsy and redbubble accounts will spring up selling these kinds of posters in the near future - if they haven't already.

Conclusion, sort of, and some caveats

So my great experiment is at an end. The takeaway is that these AIs cannot replace great illustrators nor great illustration where there are specific requirements. Basically, if it requires a proper brief and real conceptual thinking the AI is no substitute for great illustrators. If on the other hand you just want to make some cool artwork for halloween then it's brilliant fun and really very good and rewarding.

For anyone who has made it this far down and was still wondering what my cover illustration is: it's my unadulterated but cropped chosen result from the four options the Mindjourney AI offered up to me when I typed in 'W5 graphic design agency illustration experiment'.

Copyright notice: all Mindjourney illustrations are covered by a Creative Commons License.

At this point I need to add that other AIs are available, they will work slightly differently and give you different results. This experiment was solely carried out in Midjourney.

I'd love to hear people's opinions on this and please feel free to share out the article so we can have a conversation. If you know more than me on this subject - especially on the copyright issues - I'd love to hear from you and please feel free to correct me if I have got completely the wrong end of the stick on anything.

Some links for you