2 years later, Midjourney still hasn’t taken over my job as an illustrator and storyboard artist
I was right so far, but why exactly? A deep analysis and comparison of human vs AI art skills. And a test for you to try, if you still don’t believe me ;)
A lot of people feared AI will replace artists. I feared too, as an artist. But as soon as I tested the AI, I understood I was “safe”. I even wrote an article two years ago about it, that got some “success”: Why Dall-E will not steal my job.
Around me, most other artists were not so worried. It’s generally people not working as artists who jumped to crazy conclusions. And I think there are two main reasons for that misunderstanding:
The first reason is actually pretty simple: non-artists overestimate the time it takes for a professional artist to create a simple art piece (like a storyboard frame), and (vastly) underestimate the time it takes to get a decent art piece out of a generative AI model. More on that after.
The second reason is that non artists have never really tried to create something for a real client with an AI model. We artists have obviously tried (to see if we could go faster, cheat our way out of working, or simply see if we will get replaced). And we all had to revert back (thankfully), to really, hum, making art. That’s the most critical point. I’m going to expand more about it now.
Making real “new” things is impossible with AI models
I know that’s the most controversial statement. And you will test my words later in this article. But as an artist who has tried those models for a few years now, I can absolutely tell you with certainty that AI models can’t really “create” like an human, EVEN if they really, really, look like they do to non-artists (and even to some artists who haven’t fully tried it yet).
You see, for people outside of the “artistic jobs” (and even for us sometimes to be fully honest), it can be very difficult to see the nuanced differences between creation and remixing. We can only see it when we are actually trying to create something original with the AI (when we don’t have the choice). In that regard, I am thankful to AI, not only because it’s useful for some small things in my workflow, but also because it allowed me to understand better the concept of creation itself.
Let’s give an example that everybody can understand better. If I create an art piece representing a Disney princess in the style of Tim Burton, I really look like I am actually creating something new. After all, that image/concept didn’t exist before, and now it exists. But in reality, I am remixing existing things. And to be fair, to a certain extent, all art is a little bit remixing what came before. But, it’s a little bit more complicated. Let me go further before judging.
If you ask an AI model to remix, it will do it reasonably well, and generally faster than a human. It’s really good for that.
That’s why the successful AI art pieces online generally look like the crossover of familiar concepts, like Marvel superheroes as babies, Star Wars characters in Wes Anderson style, Disney characters as zombies, Countries as knights, Religions as knights, knights in the desert, futuristic knights, superheroes dressed in old renaissance style etc…
That’s what it’s made for, and it produces actually fun and enjoyable images that way. It still takes time to get good images like that, but it does it well.
The problem arises as soon as you don’t remix clear existing things. Or you don’t use an existing (and famous) style. Ask it to make a generic character in Pixar style, mostly fine. Ask it a very precise character (not random), that looks like something you or your client have in mind, that wasn’t already done before, and the results are just impossible to get. Like really. I tried.
Still don’t believe me don’t you? I see… Then don’t take my words for it, just try it for your yourself :) Here is a test for you to see:
The DIY test to see if I’m crazy
Try to reproduce this storyboard I made for a client (or even just one or two frames as a test if you want), with any AI tool you like, without (obviously) putting my images in the software (that would be absurd, the idea is to try to create something “new” without an actual exact reference, the point of my job). I trust you will try for real as there is nothing to win haha, and it’s for science ;). You can use other images found online as references though, of course, as long as it’s not the ones below. I’m going to give you the same “rules” I got from my client underneath (like a simplified script if you will):
The rules:
You need 6 images (or one or two if you don’t have the patience, but ideally six). It’s mobile format (vertical). The two characters are two young men (two travel influencers), one is white, the other one is black. They have shorts, t-shirts, and a backpack for the white guy. They need to remain the same (or relatively consistent) in all the frames, obviously. They are live streaming while in front of some beautiful landscape and old maya ruins (like the Machu Pichu kind of ruins). In the first two images, the black man is holding the phone, we see his face in the foreground. White man in the background. Then they switch, white guy holds the phone. They look happy at first, then they start to freak out when the profile pictures of the people watching their livestream change to mosquitoes profile pictures (it’s for a mosquito repellent product).
You have the right to add the livestream profile pictures of people and mosquitoes, and the likes heart icons, in “post-prod” in photoshop or similar softwares if you know how to do it, if it’s easier. But there HAVE to be some indications that people are liking their livestream, and that their profile pictures are changing to mosquitoes, otherwise the all concept doesn’t work for this particular advertising.
If you manage to create something decent, please post it in the comments, and please add the time it took you. But what will happen most likely, if you don’t cheat and use my images as reference, is that you will have great difficulty to create even a single image conforming to the brief. You’ll see. Believe me, I tried.
Why exactly?
Simply because there are no other images on the all internet of two young influencers, one white and the other one black, live streaming in this exact configuration (with the black guy holding the phone in front of his face, and the white guy in the background with the Machu Pichu or similar mountain behind him). There just isn’t. It’s not super original when you think about it, I admit, but it’s not generic enough for the software. It’s too “particular”. You can find an image of a man taking a selfie, sure. You can find an image of a backpacker, sure. You can find an image of the Machu Pichu, sure. But combining the three in a way relatively similar of this storyboard is, first, very long and difficult to describe properly to the software. But also it’s too “specific”. You will get super weird results.
As you will see, while it’s pretty “easy” (still takes time and effort but it’s ok) to get a good image of a woman having a coffee, or smiling, or a man walking in the street (because it’s generic, looks like other similar images online), it’s actually extremely difficult to get an image of something too… not generic let’s say haha.
Now if you add the livestream “effects” (profile pictures, like heart icons, mosquitoes profile pics), it’s just absolutely impossible to get that exact image. You will have to at least add it “manually”. Which is okay by the way, but still requires a little bit of “graphic design work”, if not directly drawing it.
Time
Also, as you will discover, getting a decent image using those AI models takes time. You generally don’t get a good image on the first try. From my experience, and feedbacks I read online, it takes around 15 minutes for a generic image, 30 minutes for a more specific one, and up to a few hours to get something really “great” (that people like and share on social medias, like babies superheroes…).
For an artist, it takes 5 to 10 mins to get a decent sketch, 30mins to get a decent black and white frame, 1 hour to 1:30 hour to get a finished colored frame like the ones above (a little bit more if there are a lot of feedbacks). So, for a way more accurate result, you actually don’t really take THAT much more time than an AI model. The only real advantage of the AI model is the complexity of the rendering (you can get a fully rendered image way faster than what it would take for a human, but it’s generally never really necessary for professional works to get fully realistic renderings anyway…). And the workflow is generally easier if you provide a good black and white drawing made by an artist.
By the way, the best results are generally obtained by people uploading ready made images from real artists that they found online. That’s how the best AI images are made. Not from prompts directly, but from a mix of prompts and uploading and remixing images found on Pinterest.
So real artists actually still have an advantage, as they can provide those images themselves. Which leads to the last point.
Last nail in the proverbial coffin
Now to the very last nail in the proverbial coffin, something often overlooked:
We artists also can use the AI models everybody can use.
I know, it seems obvious, but nobody seems to have given real thoughts about it :)
The fact that anybody can create “art” using the AI models actually includes… us artists.
So even if the models become very good, someone who can draw, paint, use graphic design softwares, AND use the AI model, will still have an enormous advantage on someone only using the AI models.
Think about it: if you can hire someone who can use a good AI model, but can only do that, and hire someone who can use the AI model AND can draw and paint, thus being able to modify or improve at will the AI results, who would you rather hire? Who would make the best concept arts for video games instance?
Cars are getting better everyday, but everyday people and great pilots can both drive those. So the great pilots still have an advantage on the random joe in a race.
Great chess computers can be used by everyday people and chess masters. So chess masters still have an advantage in using those to learn on random players. Sure you can beat the great players with the computer, but would you hire a random joe with a computer to teach your kids chess, or would you hire a professional grand master? Go try to cheat at chess online on Lichess.org or Chess.com, you will soon get kicked out of the servers. It’s surprisingly easy to spot cheaters at chess online. Computers simply don’t play like humans. You can win a few games, then you’ll get spotted. Any intermediate player can spot the difference. Algorithms too.
More and more of my clients are sending me briefs with AI images that they created as references. But they are still contacting me because they simply can’t get the precise results they want from the softwares. It still looks too generic, and weird.
Don’t get me wrong, you can produce super cool and beautiful images from the softwares. But everybody else can. And in those everybody there are also artists. Who can modify at will those images. What makes you think it will be easy for anybody to compete, if everybody has the same AI models?
The overall level might increase (or more accurately, the speed will increase, not the level, as the level is of course limited by the data, thus by the best artists in those datas). But so will the level (or speed) of the people at the “top” of the pyramid.
Don’t hesitate to post your results in the comments, I would love to see it, and I will be happy to see if I’m missing something :)
Have a good day!
