Experimenting with Digital Art: Text-to-Image software

The other week, I saw a fellow artist post a link on Twitter about a site called NightCafe Studio where you enter in a text prompt and it generates an image (which a lot of people mistakenly called “artificial intelligence”). Curiosity finally got the better of me, so I followed the link and began to play around with the software as an experiment. You can see the portfolio of my experimenting here on my profile at NightCafe. I’ve spent almost two weeks playing with it now, which is to say I am stepping into the concept of digital art, at least on the text-to-image side of it. (It’s so much faster than drawing or painting, and right now I feel as though I have a hundred different things going.)

The good, the bad, and the “What the ****?”

The profile/portfolio I link to over on NightCafe only includes the stuff that turned out good. I started off with a simple enough prompt, “pink flower on dark green leaves” or something very similar. If you scroll down far enough you will see the result at the very bottom. I fiddled around with a couple other phrases, and deleted those results as I did not like them. I am saving all the flower ones for their own post, but then moved on to see what kind of results I could get for some classic still life subjects.

Modern style still life of red and yellow bell peppers and one cherry tomato on white plates with a blue tablecloth
Tomatoes and Peppers, digital art

Here is one that turned out nicely. I deleted probably half a dozen or more in my attempts to get a nice still life arrangement of tomatoes, peppers, garlic, and an onion – salsa still needing assembly. I figured it would be perfect to hang up at my favorite Mexican restaurant in town, especially since my Summer Jalapeno painting is there. Let’s just say the program is not good at doing an onion and totally ignored the prompt about garlic. So instead of a still life I could name “Salsa, Some Assembly Required,” I got red and yellow bell peppers and a single cherry tomato. It does look nice, although I will need to draw/paint what I actually wanted. I uploaded the digital piece to my Pixels shop and my RedBubble shop for accessories, apparel, and prints.

Classic still life subject seem to give the art generator fits. I have some truly wild examples of a big weak spot in the code – but I will save a few of those for the floral post since I was trying for an old masters’ style of fruit and flowers still life composition. I will mention the text-to-image program has some issues with what shape apples are supposed to be. It also doesn’t seem to realize that apples and pumpkins are affected by gravity like everything else.

A smashing success for text-to-image

On a whim, I decided to try a text prompt that would describe my very first pastel piece I did at the age of just seven. I have no idea if it still exists, but it has stayed in my mind all these years. The first result is still the best: here is Sunset Palms, with the Neo-Impressionist style filter applied to it.

Sunset Palms, two palm trees silhouetted against a brilliant sunset sky with simulated impasto
Sunset Palms, digital art

The software even generated simulated impasto that looks like it was applied with a palette knife. I was a little surprised to see a swathe of green in the sky, but I think it works. I put this into Microsoft Jigsaw to play with it as a jigsaw puzzle, and the colors are just fun. Art prints are at my Pixels shop (along with puzzles) and apparel and accessories are at my RedBubble shop. I can definitely see myself painting variations of this image, either in acrylic or oil paint, once I finally get a cat-free art studio where it can dry without being walked on by my cats.

The struggle for black cats

I decided I wanted to see what kind of Halloween imagery the software could generate for me, and began to enter various prompts and try different style filters. I wanted an image of a black cat’s face filling the frame, with lovely green eyes providing a stark contrast to the black fur. Instead, the program spat out this – and I have seen hairballs that look prettier.

ugly software glitch
I just call it “fugly”

Well, at least it got the green eyes part right. No matter how I tilt my head and squint at it, it just does not look like a cat as I know them. This is emphatically NOT in the public portfolio, because that would just be embarrassing. Or maybe I am just not modern enough in my art sensibilities. I kept trying with the text prompts.

Black Cat sitting, digital artwork created using text-to-image software
Black Cat Sitting, digital art

The title Black Cat Sitting is the text prompt I used, and in this specific case, the image it gave me was very nice! I love the soft focus feel to the light, and the cat sure enough looks like a real cat. When I saw this one, I began to think perhaps there is some hope for text-to-image software after all. Yeah, famous last words. Here is what the software thinks a black kitten ought to look like.

another disasterpiece from a simple prompt
a five-eyed kitten? What the ****?

Yikes. That is not public anywhere else but here on the blog. Seriously – does this have five or six eyes, and why would a supposed “artificial intelligence” think kittens look like that? I mean, I know Halloween is supposed to be all about scary imagery, but this is just so many levels of wrong. And that is about as polite as I can stand for this one. I could post a few jack o’lantern mishaps, but I think y’all get the general idea of just how spectacularly wrong the algorithm can be on some of these.

Text-to-image digital art of – ME!

One day last week, I had an “Ah-ha!” moment and decided to see what kind of software-generated digital image I could get from the text prompt describing me. Hold your laughter now. I started off with the text prompt “middle-aged woman with short blonde hair,” and ended up with this:

ugly image from what ought to be a straightforward text prompt
how I feel first thing in the morning

OK, now y’all can have a good laugh. Apparently, punctuation is important, and that hyphen in middle-aged must have messed the algorithm up good and proper. Granted, there are definitely mornings when I might feel like this, but I don’t usually look quite this rough (or so my husband says). So I tried “mature woman with short blond hair,” and the second time I got an image I am now using as my “face” here on the internet.

the internet me, from text-to-image digital art program

Now, she doesn’t look like she’s about to celebrate her 50th birthday like I am, but most of the female figures the software spits out look to be no older than 25, and most look like they could be movie stars, so an image that brings to mind Florence Henderson from The Brady Bunch probably fits the program’s parameters for mature. I like the simulated brushstrokes for it as well. Her eyes are the wrong color and she isn’t wearing glasses “thick enough to see into the future,” but she’ll be my face for the internet for now.

While the text-to-image programs used to generate digital art are far from perfect, every once in a while you get a good result. I can’t see it replacing us traditional artists any time soon, but for now it is an amusing novelty. I may just use it as a way to test composition ideas or to get inspiration. Mostly, I am just playing with it because it amuses me.

13 thoughts on “Experimenting with Digital Art: Text-to-Image software

  1. Looks like a fun toy, I guess, although obviously the novelty will fade.

    My own feeling is that artists and photographers should realize that at some point the only thing we’ll have left to offer is “authenticity”. Anything else in imagery will be available right off the shelf.

    As someone who worked in software development for many years, all I can say is that you’re right, this isn’t “artificial intelligence”. In fact there is, today, no such thing. But there are some very clever algorithms working with very large data sets at very high speeds…

    1. The novelty will likely fade, but by playing with it I can offer informed commentary on both its strengths and weaknesses. Most of the people hyping it only talk about the good parts, and never mention that there are definite weak parts of the code. More on that in posts to come.

  2. Looks like fun but, alas, not for me. I have too many irons in the pot an enjoy the act of going out and discovering images too much to allow myself to get drawn into another reason to sit in front of my computer. But I’m sure others will enjoy doing it.

  3. Intriguing, indeed! It made me think of the discussion we had earlier in the year when I created some digital paintings from my photographs and how that made you feel as an artist. Looking at these, and I recognize that they are the best of the bunch (mostly), I am thinking the same sort of thing. Why go for a perfect still life photo (or painting) of flowers in a vase when many people would be more than happy with a print of the flower images you have included above (and more on your site)? Food for thought indeed!

    1. It made me think of our prior conversation (which I link to above) as well. As for the why of trying to get that painting (or photo) of the perfect still life or bouquet of flowers … because it is a goal unto itself. Just like I can usually draw or paint faster than I can get a truly good photo, it takes quite a few days of playing with text prompts and modifiers before the nice stuff starts happening regularly. Most of these were just a matter of luck – at least, that is how it feels to me.

Leave a Reply

Your email address will not be published. Required fields are marked *