Putting AI-generated artwork to the test

by | Dec 14, 2022 | Ai art

Two real-world applications in my job

In my last blog post, I took a deep dive into the future of artwork created by artificial intelligence. My emotions ran the gamut of amazement to existential panic as I pondered my future as an artist and human.

While I don’t expect to be replaced by a robot graphic designer anytime soon, apocalyptic predictions abound on the interwebs regarding the doom of all kinds of industries. In short, pretty much anything built around entirely digital assets looks to be on the chopping block – stock photo websites, video game artwork, and content creation in visual or written form. 

I’m immensely interested in this idea because I work with stock photos a lot. My favorite websites are Envato elements, depositphotos.com, and shutterstock.com. Sometimes I use the stock images exactly as they are, but more often than not, I alter them to fit the situation, brand, or style of my creative brief. The key to using stock assets is to make them look like they’re NOT cheap generic imagery. Depending on the project, I can spend minutes or hours altering a stock photo or illustration.

Here’s an example: 

BEFORE: Stock photo of a blank book
AFTER: Art applied by me to create a mock-up

What if I could cut down the time I spend altering and editing stock images? What if I could type in exactly what I wanted and get it on the first try? And what if everyone, not just graphic designers, could do that? 

Challenge accepted! I wanted to see for myself if I could use AI-generated assets in some REAL projects. Here are two quick case studies. 

Example 1: Fall Festival Graphic 

I was tasked with creating a digital fall festival graphic for a local historical attraction. First, I tried to see if Dall-E 2 could just do my job in one fell swoop by creating a completed graphic. I typed “fall festival poster” into the prompt.

Whoa, interesting! “Fatal Fail” isn’t going to work at all. Obviously, the AI can’t write words and its attempts to do so are hilarious. Beyond that, however, some of these concepts struck me as workable. I really liked some of the color palettes, compositions, and illustration styles. With a little (or a lot) of polish, I could definitely “cinderella” some of these abominations into a finished product.

But something else was bothering me. I was kind of puzzled – questioning my own judgment. Was I attracted to some of these concepts because I really liked them? Or was I grading them on a curve like I do when my kindergartener draws disfigured but recognizable portraits of family members? This raises bigger questions, but that’s another blog post.

Stick to your process, Barton

I wasted 2 hours trying all sorts of things, but eventually realized I needed to literally go back to the drawing board. I started by developing concept sketches on paper.

Here’s the rough sketch I landed on.

With this sketch, I have a roadmap to make the final artwork.

Now that I know where I’m heading, it’s time to gather the building blocks. It’s kind of like building anything, you need raw materials. In this case, I needed a pumpkin for the foreground and a marsh scene for the background.

Normally, this is where I go to the previously mentioned stock websites. This time, however, I headed back to my AI co-worker and cast about for a marsh scene: 

For this batch, I instructed the AI to create an “oil painted” look.
For this batch, I requested a flatter more illustrated vector-style graphic.

Pretty impressive results. But you don’t see cattails in a Lowcountry marsh so I fiddled with the prompt (i.e. typed in different words). Here’s the one I eventually settled on:

Scary accurate.

Then, I searched for a pumpkin:

I opted for a simple digital version (top row, second image). Then, I worked my designer magic (e.g. re-colored, simplified, added text) until I landed on this:

Admittedly, I significantly altered the original assets and it took longer than usual, but we got there! And the process wasn’t that different from how I usually work. In the right situations, I can see AI-generate stock assets being a regular thing for me – especially at the conceptual stage of ideation.

Example 2: Map and Brochure 

The next example project involved a printed brochure I designed for a client. Here’s the flat artwork that I created.

Not very dynamic when displayed in this manner.

I wanted a photo of the final printed product, but I don’t have a physical brochure. So I decided to make a mock-up in Photoshop. I could go to a stock photo website and find a good mock-up. Instead, I prompted one.

I requested a blank bi-fold brochure mock-up from the “model”. The model is another way of saying AI programming. Here are the results:

The wonky ones really make me chuckle.

I grabbed the top right, did a little correcting in photoshop, and then applied the flat map artwork.

What do you think? Not bad, huh? It’s got some imperfections and therefore feels a little more “real” – but that wasn’t what I was going for exactly. I also like that it is 100% unique, you won’t find this mock-up anywhere else.

But was it worth my time? No, not in this second case. And not in the third, fourth, or fifth! I tried several different experiments of varying complexity and the AI generally failed. Perhaps it could succeed with better instruction, but I don’t have that kind of time at the moment. In business, you need a sure bet.

The limitations to AI art are still very severe. That said, the pace of improvement is mind-blowing. I do expect these tools to become viable options for businesses like mine in the very near future. I’ll be watching and learning. I’m even excited about seeing how things change.

There will be disruption. Time will tell how much.

Have a great week,