In order to maintain full transparency and ensure the quality of our art, here’s a breakdown of our creation process.
Remember, we use AI art to share the message of Jesus Christ. So the initial image(s) are created by AI, and then edited/altered by our graphic designers. They are prompted and edited by our prompt engineers. Here’s the breakdown as explained by Chas Hathaway (our CPO, chief prompt officer 😊):
When I come up with an idea for a piece of art, I’ll often get online to see what’s been done. I don’t want to duplicate someone else’s work too closely, and I also want to understand what people are familiar with. Mostly, however, I want to get ideas of how the story can be portrayed that will be unique, beautiful, interesting, and most of all, emotionally impactful or faith promoting.
Just to give you an idea how this all works, let’s walk through the creation of a piece. If I ask for an image of Jesus with a child, that comes fairly easy to AI (since it’s been done SO MUCH over the centuries, and therefore informs the databases from which AIs are trained). With an easy subject like that, if all goes well and a few prompts are all that is needed, I can skip straight to the upscaling step. I’m choosing a tough subject so you can see how in-depth this process can become. And I’ve decided to do this in real-time, not explaining what I was thinking at the time, but actually come up with it WHILE writing this process. In other words, it could be a complete failure. But even failure is part of the process, right? If that happens, we’ll change subjects and try another, since that’s really what I do.
So… ready? Here we go!
Step 1: Choosing a Subject
I’ve always wanted to get a good painting of Jesus Christ healing the paralytic man who is lowered through the roof of a house Jesus is in. I’ve actually attempted it briefly several times before, with very mixed results—mostly unusable ones (I say mostly, because I think at least one result I ended up using for an unrelated subject, since it looked more like Jesus healing a blind man, or something like that). So here’s the story, as related in Mark 2:
And again he entered into Capernaum after some days; and it was noised that he was in the house.
And straightway many were gathered together, insomuch that there was no room to receive them, no, not so much as about the door: and he preached the word unto them.
And they come unto him, bringing one sick of the palsy, which was borne of four.
And when they could not come nigh unto him for the press, they uncovered the roof where he was: and when they had broken it up, they let down the bed wherein the sick of the palsy lay.
When Jesus saw their faith, he said unto the sick of the palsy, Son, thy sins be forgiven thee.
But there was certain of the scribes sitting there, and reasoning in their hearts,
Why doth this man thus speak blasphemies? who can forgive sins but God only?
And immediately when Jesus perceived in his spirit that they so reasoned within themselves, he said unto them, Why reason ye these things in your hearts?
Whether is it easier to say to the sick of the palsy, Thy sins be forgiven thee; or to say, Arise, and take up thy bed, and walk?
But that ye may know that the Son of man hath power on earth to forgive sins, (he saith to the sick of the palsy,)
I say unto thee, Arise, and take up thy bed, and go thy way into thine house.
And immediately he arose, took up the bed, and went forth before them all; insomuch that they were all amazed, and glorified God, saying, We never saw it on this fashion.
Mark 2:1-12
Now, there are many parts of this story that we could portray. We could do an image of the people carrying the man up to the rooftop. We could have him in the midst of being lowered. We could do the moment Jesus is using His miraculous power to heal the man. We could get Jesus rebuking the Jewish leaders who think He’s being blasphemous. We could even do the moment the healed man stands and takes his bed. So how should I pick the moment to portray?
I think I’ll start by seeing what’s been done. So let’s do a Google image search to get a quick, rough idea:
And here’s a similar Pinterest search:
Though each of the images (the ones that are actually about the subject, anyway) seem to be conveying the situation well, I don’t see any portraying the surprise that must have been on the faces of the disciples as they realized what was happening. The show, “The Chosen,” portrays well the moment when Jesus recognizes that an act of profound faith is taking place—and in the midst of a bit of crowded chaos, too.
https://www.youtube.com/watch?v=dlBOmQ1PaMY
But even here, there’s not a single frame that portrays both the surprise of the disciples AND Jesus’s recognition of faith. They didn’t have to do it in a single frame, since this was video. But I’m going for a painting, so it’ll have to be in a single frame. No guarantees, but let’s see if I can come up with a unique painting of that moment of surprise and a recognition of profound faith. But we’ll have to be flexible, because if there’s one thing I’ve learned about AI-based art, it’s that it’s an incredibly imprecise science.
Step 2: Choosing a Style
There are so many great artistic styles. But I have discovered a pattern that the simpler the scene, the simpler the style can be. Jesus with a child can be a very simple scene. An incomplete pastel watercolor style would work great for that. This is NOT a simple scene. It’s rather complex, and I really want to do it justice, if at all possible. So what art styles lend well to complex scenes?
When I imagine the surprise on the disciples faces, I think of that fun baroque style with unique faces on each person. So I Googled “baroque art” until I found an example of this:
It’s almost a little Norman Rockwell-ish. And while I know some of the artistic terms, I’m no pro at them. So I’m going to get on ChatGPT and ask how it would describe this painting.
And here’s ChatGPT’s response (italicized):
This painting, “The Musicians’ Brawl” by Georges de La Tour, belongs to the Baroque period, which flourished in Europe from the early 17th to mid-18th century. The Baroque style is characterized by dramatic use of light and shadow, intense emotions, and intricate details.
Genre
- Genre Painting: This painting falls under the genre of “genre painting,” which depicts scenes from everyday life, often with a narrative element. In this case, it captures a moment of conflict among musicians.
Style
- Naturalism: The painting exhibits naturalistic detail, portraying the figures and their emotions with lifelike precision.
- Tenebrism: Georges de La Tour often used tenebrism, a style marked by the use of stark contrasts between light and dark to achieve a sense of volume and three-dimensionality. This technique is evident in the dramatic lighting of the figures against a dark background.
Artistic Techniques
- Chiaroscuro: This technique involves the use of strong contrasts between light and dark to give the illusion of depth and volume. La Tour’s mastery of chiaroscuro is apparent in the way he illuminates the faces and hands of the figures, creating a sense of three-dimensionality.
- Detailed Texturing: The artist meticulously renders textures, such as the rough fabric of the clothing, the smoothness of the musical instruments, and the aged skin of the musicians, enhancing the realism of the scene.
- Expressive Faces and Gestures: The painting captures a moment of high emotion and conflict, with each figure’s expression and body language contributing to the narrative. The detailed portrayal of their faces and gestures adds to the dramatic intensity of the scene.
Overall, Georges de La Tour’s “The Musicians’ Brawl” exemplifies Baroque artistry through its dramatic use of light and shadow, naturalistic detail, and vivid depiction of human emotion and interaction.
So, I’ve never heard the term “genre painting,” or “tenebrism.” But I know I like paintings using chiaroscuro, detailed texturing, and expressive faces and gestures. Maybe those will help us come up with a good initial prompt.
Let’s try using this style.
Step 3: The Initial Prompt
Here’s where we start putting AI to work. I will say, how you format your prompt may depend somewhat on which AI you’re using. I’ve found ChatGPT to usually be the best starting place for me. If it fails me repeatedly, I usually go to Bing Create, which uses Dall-E 3. But let’s give ChatGPT the initial prompt. Let’s try this:
“Please create an original, unique, widescreen painting from the biblical story of Jesus healing the paralytic man. Have it take place in a simple, first-century middle-eastern home where Jesus is speaking to a group of people. Portray the man being lowered through a gap in the ceiling on a hammock-like fabric structure. The man is in poor clothing, and his legs are limp and paralyzed. The hammock-like structure is being lowered with ropes held by several people on the roof. Create this as a genre painting in the tenebrism style. Give the people in the house expressive faces and gestures. Employ chiaroscuro technique, along with detailed texturing and naturalistic, lifelike precision. Give Jesus an expression of pleased recognition, but everyone else varying expressions of surprise, confusion, or anger.”
Now, this is a pretty detailed request. It wouldn’t even be possible in Bing Create because of it’s length, but let’s give it a go on ChatGPT and see what happens. Here’s GPT’s first result:
I’ve often told my wife that we could make an entire blog of AI image fails, because they are so abundant. And while there are obviously so many problems with this first image, I’ll be honest, I’m impressed! The biggest problem, obviously, is… why is JESUS in the bed being lowered?
Then, of course, it forgot to do it widescreen (I could talk a lot about why I request this, but it basically comes down to more bang for your buck as far as the size of the image). But since that is in the initial prompt, and we don’t know why it put Jesus in the bed, I’m simply going to have it try again. Mind you, I may still use parts of this initial image, since much of the art isn’t bad.
So now I’ll type in GPT, “Again.” Here’s the next result:
About 10% of the type, GPT likes to sent two images at once, and it’s rare that these turn out any good at all. It’s as if it’s splitting the quality in half on each. So as I usually do, I’ll just click the little arrow circles button to “regenerate” (all the while trying to ignore the terrifying result in the first image, and the laughable ceiling-pole-heaver in the second image). New result:
First off, I’ll DEFINITELY click “Better.”
We can still find many flaws in this image, but it’s kind of working for a good part. So I’m also going to thumbs-up this image to hint that GPT is starting to get on the right trail. I have yet to determine if the AI intuitively knows its not supposed to be Jesus in the hammock, because the man has a halo, while the man on the ground doesn’t. Let’s just see what it does if we type, “Again.”
Whether this is an improvement or just a variation is hard to say. At least it’s remembering to widescreen it. You’ll notice I didn’t address that issue to the AI. It’s hard to know which mistakes are digital oopsies, which are technical challenges, and which are prompt problems until you see a few results. Either problems immediately go away, or they don’t. That one went away, so I don’t have to address it. The Jesus being in the bed issue is also reducing, so that’s probably okay, too. I think the AI knows wheel chairs weren’t around in those days, and this is the first time it did that, so I’ll let that slide for now. The perspective issue with the man in the blanket (which should go behind his shoulders instead of in front of them) is a form of AI “hallucination,” albeit a mild one. Such hallucinations don’t usually perpetuate, so I’m going to thumb this image up and have it try again.
We’re having repeated hole-in-the-ceiling issues, but it’s changing every time, so I think it’s trying to get it right. Again:
I kind of want to like this one. But our original intent is lost (the surprise, faith, etc). Let’s see if that’s a fluke.
Step 4: Tweaking the Prompt
Okay, I feel like we’ve seen enough images to get an idea what the AI thinks its supposed to be doing. So let’s address those things. Systematic concerns: The style is feeling a little simplistic. While the very first image had that contrasty, detailed baroque style, these later ones are feeling like images from a children’s “Bible Stories” book. We want to create a work of fine art.
And the wheelchair is back.
The people are looking rather European—an issue that requires CONSTANT vigilance in order to overcome, if it’s even possible.
So let’s prompt GPT the following:
“Please keep the fine-art details of the baroque genre, using tenebrism style. Give the people in the house expressive faces and gestures. Employ chiaroscuro technique, along with detailed texturing and naturalistic, lifelike precision. Give Jesus an expression of pleased recognition, but everyone else varying expressions of surprise, confusion, or anger.” (you might notice that I’ve copied and pasted from our original prompt everything after the first sentence of this prompt).
A little terrifying, but let’s give it the benefit of the doubt:
“Again.”
Paralytics falling from the ceiling now?
Perhaps you can see why this painting has been a struggle so far. Oh, and I forgot to address the wheelchair. While some people encourage negative prompts with AI, sometimes these can confuse the system, and you’ll find that everyone has a wheelchair. So I try to start using a roundabout wording.
“Please keep the items in the painting consistent with 1st century Jerusalem, and have everyone be of middle-eastern descent, with olive-toned skin and black hair.”
Ancient wooden wheelchair. Nice. And we’re now preaching in an infirmary. Let’s see if it’s a fluke.
“Again.”
Okay, now we’re talking. I feel like we’ve got something we can work with here. Yes, it’s going to need some work, but it makes a good foundational image. So let’s download this and get started in Photoshop!
Step 5: Editing
Now we’re not done with ChatGPT, but we’ve now got an image to work with. We’ll start by getting that image as close as we can to a final product.
Let’s start by fixing the ceiling/roof. There shouldn’t be bars over it. There are lots of ways to remove them, but let’s start by trying Photoshop’s generative AI. I run the lasso tool over the unwanted bars, and in the GAI (Generative AI) window, I’ll type, “Remove.”
The result is three options, all of which still have bars, so I’ll just click “Generate” again to get three new options. And this time, there is ONE of them that doesn’t have bars, and it’s not bad.
Now let’s deal with the fact that his hammock is (at present) hanging from the ceiling, rather than being lowered through the ceiling. Again, there are lots of ways we might deal with this. I have a feeling GAI won’t do it on its own, but let’s see what happens. We’ll lasso the rope on the right and tell the GAI window, “Rope coming down through hole in ceiling.”
The result was better than I expected. Of the three options (six, once I clicked “Generate” again), this was my favorite so far:
I mean, sure, it’s not coming out of the hole, but the rope isn’t bad.
So now I’ll lasso some of the hole to see if I can simple copy and paste enough of it over to make sure the rope is coming from the hole, not the ceiling. And anytime you’re selecting something to copy, paste, or move, it’s always a good idea to feather the selection. I feathered mine 3, with the idea that I’ll just erase whatever I need to.
So the hole isn’t too bad, but I’ll need to cut out the rafter and ropes. I’ll use the lasso tool for that.
It’s not great, but I’ll use the eraser tool and smudge tool to clean it up a bit…
That’s a little better. Now I may come back and clean that up more, but at least it’s plausibly in the right place, so it’s good for now. So we’ll now flatten the image. That’s not essential or anything, it just helps me feel like I’m decluttering as I go.
Now I’ll fix his paste-white foot.
Grrr… while trying to catch a quick image of his foot, I bumped some wrong key combination, and my whole image now looks like this for some reason:
Since Undoing didn’t do anything, I figured it must be a “View” thing or something, but after several vain attempts to figure it out, I Googled the problem. Turns out instead of Command-shift-4 (selected screenshot), I did Command-4, which changed the “channel” of the image, whatever that means. But it became normal again when I hit Command-2. So, once again, here’s the foot.
I mean, part of me says, “He’s paralytic, maybe this disease or gangrene or whatever is the cause,” but this random grayness is a classic AI art glitch. If it’s gangrenous or whatever, that ought to be intensional. So we’ll lasso the off-colored portion, feather it about 5 or so, and then mess with the hue/saturation, and maybe colorize it. I also Gaussion blurred it a bit to make it look a little better. Result:
And in classic AI fashion, the number and configuration of toes is just a little wonky, so I’ll see if GAI has a better idea. Lassoing toes, and…
Though not ideal, it took a few tries to get it to do much. Photoshop seemed to think we were trying to edit something inappropriate that’s against their guidelines. Frankly, if it thought that, I don’t want to see what it would have produced if those guidelines hadn’t been there. Now I can just use the lasso tool to “draw in” toe grooves and toe shading. The result…
Of all the toes I’ve ever done in my whole life, these were… among them. It’s work. The more I look at the foot, the more I see the heal looking stranger. But I’m not sure there actually is anything wrong with it, so let’s move on and we’ll come back if it looks like it needs it later. Let’s fix the other rafter, to make sure it’s coming out of the roof, instead of the ceiling.
For this, I’ll try lassoing the ropes to free transform them just a touch.
Here you’re seeing the new on top of the old, and it looks like it’s getting us in the right direction. Now let’s remove the old. A little lassoing, a little GAI work to fill in the background and soften edges, as well as to remove the extra strip of cloth extending from his foot, and…
Now we have the ropes being lowever from the hole in the roof.
Next I got rid off the extra hanging ropes around the house. No one’s hanging themselves today, thank you. Lasso ropes, GAI “remove”, and Bob’s your uncle (whatever that actually means… anyway, done!)
Now let’s give the guy a little dignity. While it may be safe to assume that he has some kind of waist cover, viewers won’t have to even question it if we simply give him some kind of clothing. I could GAI this, which with clothing can be quite interesting, and sometimes effective. I think it’ll be faster just to borrow something from another part of the picture. How about from the guy on the far right who’s already looking a little skeptical about the whole scenario. I’ll lasso a bit of his clothes in the shape of the needed clothing on the paralytic man.
Okay, so not all that convincing… yet. Let’s try employing some principles of shading, maybe trim it up a bit, and see if it can become passible…
I had to borrow some shoulders from other guys in the image, but I think it will do for now. Now, as eager as I am to fix that hand (and face, oh my garbled), I’m going to wait on that still and HOPE I don’t forget.
Now, Jesus was accepting of everyone, and some level of skin showing is natural, of course, especially among the struggling dirt-poor class of people. But this entire group is probably ancient Jews. They never would have gone around in public completely shirtless, so I’m going to throw some tops or robes on the other shirtless guys in the room. This time I’ll start with generative AI, just to see how it goes…
You know what? I’m impressed with the very first result. I think I’ll keep it.
Okay, so what next? Anything else weird in the room? Hmmm… well, I don’t know how ancient houses normally went back then, but I imagine they usually had a closing door. So let’s turn our public arched entrance into a private door if we can. We’ll again start with GAI to see if it will make it easy. I’ll select the opening and the ladder (I mean, why is that there anyway?) and type “ancient door” in the generative window. After many cathedral doors, random windows, and a few Catholic church doors, I changed the GAI prompt to, “Simple ancient wooden door.”
This didn’t help. I won’t bore you with all the failed “doors”, but I’ll show you a snapshot of the thumbnails of a few of them it suggested…
So instead, I made the prompt blank to see what the AI would come up with on its own. No go. Just more arches that look much like the original.
So I went back to one of the earlier GAI results and found the one that most closely resembled what I’m looking for. We’ll try this and go from there:
Now let’s try GAI trimming the top off…
It took a few tries, but I think this will do.
Now let’s clean up the room from anything that doesn’t make sense. I’m okay with clutter, but anything nonsensical could be a distraction. I’ll lasso and GAI “remove” for everything we don’t want.
Ah, if only I could clean up my house by selecting and typing, “Remove!”
Okay, now let’s do a digits test. Check all fingers and toes, make sure there is either sufficient obscurity or the right number/shape of fingers. As predicted, the hammock man was the hardest. This one of the early generations that was actually recognizable as a hand.
In the end, I think the hands and feet came out alright.
Now, while not at ALL confident that I’ll be able to pull this one off, I’d like to see if I can get a little bit of an image of the people on the roof helping the man. For this, I’m going back to ChatGPT, since Photoshop’s GAI really isn’t very good at creating convincing things that aren’t there at all. The way I’ll do this is to continue the same thread I was doing, with the prompt, “make the people on the roof that are lowering the man down visible through the hole in the ceiling.”
For the sake of time, and to keep you from getting bored, I won’t show all the versions of this. Mostly it’s me typing, “Again.” “Again.” “Again.” to try to get it to show me something I can use.
But the result was repeatedly the same kind of images with NO people on the roof at all. So I got more aggressive and said, “Have the roof above and the hole in the ceiling full of people also.”
But however many times I repeated it, corrected it, instructed it, it wouldn’t put people on the roof. (Meanwhile, the wheelchair came back with a vengeance.) So Now I’ll try another tactic.
“Have people pulling on the ropes of the man’s hammock/bed”
My favorite fail, I call, “Ancient zipline slide”
“Again”
“Again”
“Again”
Someone on the roof finally! Even Jesus is surprised. Too bad they look like scarecrows.
I can’t even get the other people to touch the ropes. Gratefully, the guy at least grabs his own ropes here and there. We might be able to use this…
What I’m thinking is… what if I take that guys arms, and superimpose them as the arms of the people on the roof above?
Oh, well.
Maybe instead of having arms holding them, we’ll just sneak some unwitting characters up there from some of the many images we’ve gotten so far.
You know, it almost works. But try as I might, I couldn’t get a convincing background against the roof-dwellers. Plus, even if I did, the faces are so badly garbled that I’m not sure it would be a good idea anyway. Oh, and I forgot to take my ridiculous arms out.
So let’s jump to the next step before continuing our editing. I think that will help.
Step 6: Rescaling
This part’s simple. In Photoshop, I go to Image>Rescale, and choose a bigger size. Photoshop’s tech makes it as good and clean a resize as you can reasonably expect so far. I like to do a DPI of 300, and the smallest size at no smaller than 6000. That will make this image 6000×10500. Oh, and let’s set the noise reduction to 60%. This is a rather noisy image.
Step 7: Continue Editing
The image will look exactly the same from the computer at this point, just bigger, with maybe a touch less noise. Now let’s fix up some garbled faces. My experience is, if I fix one garbled face, I’ve got to fix them all. To do this, we’ll head back over to ChatGPT and start requesting closeups. I don’t do this on all the images I make, just on the ones that have several far-off faces or several garbled faces. In our case, you can bet we’ll be doing a lot of it.
We’ll do Jesus first, since He’s the most important subject next to the paralytic man. I’ll prompt, “Let’s try a closeup of Jesus’s face, looking up at the man.”
Here’s our first result:
Not quite what I’m looking for, but I think you can see now why I would use this method. Fine detail comes back in quickly. Now let’s fix the race, since European Jesus has returned. But instead of saying middle-eastern, I’m going to get even more specific. There are a few races living today that probably closest resemble what Jesus and the Jews of His time would have looked like. My favorite (in terms of the results AI provides) is Palestinian, so let’s do that.
“Make him Palestinian, and give him a simple hood”
Not bad, but let’s keep going. Now I’ll just keep typing, “Again,” and “Try again” until I get something I really like…
Okay, I found one that might work… and when I copy and paste it into our image, it’s comes to this size:
Which is obviously a little too big, but relatively speaking, it’s pretty close. So I’ll resize the head, tinker with the surrounding area, and make it mesh as a single image.
A little resizing, hue-tinkering, smudging, and erasing and… wallah!
One head down, only like 30 or whatever to go.
Sigh…
So we’ll prompt ChatGPT on the same thread, “Now let’s do a closeup of one of the other people in the room, looking up with wonder and confusion”
Okay, the position is okay, but I feel like we’re losing the style a bit. But, to give it the benefit of the doubt—
“Again”
Yeah, time to remind it of the style we want. Time to use another cut and paste from the original prompt.
“Do it as a genre painting in the tenebrism style. Give the people in the house expressive faces and gestures. Employ chiaroscuro technique, along with detailed texturing and naturalistic, lifelike precision.”
Yeah, there we go. We can work with that. We’ll now do the same thing we did with Jesus’s face/head.
Perhaps you can see where I’m going with this now. This isn’t terribly difficult to do, but it does take a little Photoshop saavy—or else Googling skills to find out how to do things in Photoshop. Most of all it just takes time. Not exactly the “generate and post” people sometimes associate with AI art. But AI is a great tool for creating the building blocks for a great painting.
So without boring you with the re-building of every face in the original image, I’ll repeat this process of finding a good face and Photoshopping it to the body in the image.
And as tempted as I am to leave Mr. “killer hangover” in the image, I’ll have to fix him, too.
So, without further adieu, I’ll fix ALL the faces (and possibly delete some, add some, etc), I’ll do all that and get back with you.
Many Fixed Faces Later
Took a couple days, but drumroll please…
So, I still need to adjust the coloring/lighting a bit on the faces, but we’ve come a long way. I did with each face what I did with Jesus’s face, getting individual faces from ChatGPT, feathered, trimmed, flipped, and did whatever I needed to make them look decent. And even hangover guy, while no longer the same face, is still our subtle comic relief.
Now let’s touch up the lighting on the faces that need it.
Okay, I feel a little better about this.
Yay! Our editing is done! Time for the next step…
Step 8: Stylizing
I haven’t been flattening the image in awhile, since I knew I’d be adjusting the lighting/hue of the individual faces, so I’m going to do that now. It won’t look any different, but this step is essential at this point, or you’ll end up just stylizing a single layer. The whole point of this step is to homogenize everything. If you look close at the image now, (hangover dude, now goofy dude, is a great example) you’ll see that there is a difference of quality/style between his face and his body. Obviously this would be distracting to the viewer if it stayed this way. Stylizing the entire image helps mitigate this problem.
So we’ll go to Filters>Filter Gallery. In Filter Gallery, you can see a preview of what it will look like based on the different filters.
Now to give an idea of what different filters do to the image, I’ll snap a few screenshots of different filters. You’ll quickly see that some work better than others.
I’ve only put a few of the filter previews here, because there are quite a few. But as you can see, some filters are better than others, and as you can imagine, different filters are going to be useful for different kinds of images. For most of the AI paintings I do, Dry Brush, Palette Knife, and Watercolor are the most common, since they seem to be most true to the original.
I like the dry brush best for this one, so I’ll choose that one. Then I’ll adjust the settings on the brush to allow most detail through: Brush size 1, Brush detail 10, and Texture 1.
Tadah! Now, I could apply more filters, such as giving it a canvas texture, but I don’t think I’ll do that here. In other words, the image is done. Which means on to the final step.
Step 9: Exporting the Image
Now I wish this step were as simple as “Save,” but it’s a little more than that. First I’ll export the file as a full-size, full-quality PNG file. Again, nothing really changes the look of the image from this point on, but without these last steps, I still wouldn’t be able to share the image with everyone.
So I’ll save the PNG file, and choose the title for the piece, “Healing the Paralytic.”
Now I’ll go find that file on my computer and open it in GIMP.
Yes, I could do this last step in Photoshop, but Photoshop only has 7 JPG quality settings, whereas GIMP has 100. Why does this matter? Well, it probably doesn’t matter a ton, but here’s the issue: Etsy will only allow you to sell images that are 20 MB or less. When I first export this PNG, it is 52 MB. That’s not going to work.
So I’ve opened it in Gimp. Now, there are a few ways I can bring the file size down. I could shrink the image. That would shrink the file size, but that would also shrink the printing size, limiting what customers can do with the image. I could simplify or re-stylify the image with a style such as “Cutout” or “Greyscale”, but for obvious reasons, I don’t want to do that.
The simplest, most effective method I’ve found is to export it as a JPG and bring the resolution down a touch. I never want to drop it more than 75%, but I aim to keep the resolution at 90% or higher. So what I do is, in Gimp, export the image at, say, 95%.
Then I navigate to the file and pull up the info tab to see the file size.
Look at that! We have some room to work with. It’s only about 13 MB, so we can up the resolution a bit. So I go back to GIMP and export again, but this time, setting the resolution to 97%. And yes, I want to replace the previous image.
This time it came out 17.8 MB. Not bad! I could use that. But JUST IN CASE I can bump it higher, I keep going. 98%.
BOOM! We’ve got it coming to 20.7 MB. But didn’t you say Etsy only takes it up to 20 MB? Yes, because that’s what Etsy says. But whether they know it or not, anything under 20.9 MB will be accepted. How do I know this? Well, I’ve uploaded a LOT of images, and that seems to be the true breaking point.
Anyway, so we have the largest, highest quality image Etsy will take.
Step 10: Thumbnail version
Next step is to make a thumbnail. I simply scale it down to a smaller image, usually about 1138 X 650. Now, this isn’t technically one of the steps in the artistic process, per se, but it is one that provides incentive for people to actually BUY our AI generated, human edited images. See, what you’re seeing on Etsy, and on our social media, is actually just the thumbnail. If I upload the full-size image, not only does it take forever to post everywhere, but it means that in some of the places, someone could download the image at full quality without paying for it.
Now, we’re not Nazi about it. We’d never get after someone for taking our images and running off with it. We WANT the images shared. We WANT them to spread. But we do hope that if they’re planning to print them out huge, or use them in a super HD movie or something, that they’ll buy the full-quality version. We would like to make a living doing this so we have the time to spend to keep doing it.
And how different do the thumbnail and full quality look? That’s tricky to show on a computer (which is inherently a lower quality image) unless we zoom in. So let’s look at the zoomed in image of Jesus’s face in this picture in the thumbnail AND the full quality version.
So there you go! There’s our process of making AI generated art. It’s not effortless. This image took about 2.5 days to make, which admittedly is MUCH faster than if I had painted it myself. But the work put in is worth it because I can put out many more works of art.