Fundamentals of AI Art Generation

[Macavity116]

When you say you're telling the software to "redraw" are you re-using the previous image as a guidance input, adjusting the previous prompt, or both? (Forgive my seeming ignorance, I am making my first serious attempt at using an AI image generator right now and am just kinda stumbling through)

I discovered a fatal flaw. The software doesn't know what a hammerhead shark is, and therefore, it can't draw an alien reptile with a hammerhead head.

Fascinating. Maybe mixing and matching animal parts in general is a stumbling block for the machine?

 

[CBR JGWRR]

When you say you're telling the software to "redraw" are you re-using the previous image as a guidance input, adjusting the previous prompt, or both? (Forgive my seeming ignorance, I am making my first serious attempt at using an AI image generator right now and am just kinda stumbling through)

When I used redraw in that last post, what I'm referring to is that the software I use has the option to "inpaint" an edit; you draw a mask over the area you want to edit, describe the edit you want to make, then tell it to generate. I also add copying the original first, just in case the result is only a candidate for retention rather than perfect or junk. (junk you undo and tell it to generate again)

Fascinating. Maybe mixing and matching animal parts in general is a stumbling block for the machine?

Well... It's hit and miss. On this occasion I put "Hammerhead shark" in to test whether the software knew what one was, and in eight images I got back seven Great Whites with oversized jaws, six in the exact same pose and having it's shark bursting from the water in the same pose used in the other images, and one proportionally correct Great White.

 

[Macavity116]

Macavity116's attempt at AI images

This is actually my second attempt at using AI images in one of my Stellaris stories. A little less than a year ago, I tried to create some AI-generated artwork for Year of Hell. All of those images turned out to be so bad that they were never used. I followed the examples laid out by @Chac1 and @CBR JGWRR to make this second attempt work.

I have not committed to using AI images in my next Stellaris story yet, but this outing has been much more successful. Unsung Paragons of the Forgotten War is probably months away from its debut on the forums, but right now it's shaping up to be a good opportunity for me to take AI art for a test run.

I used Leonardo AI to create all of the images below. I started out by creating portraits of the main protagonists and got three that I wanted to keep:


Main Protagonist Ino Esperanta:

For Ino, I wrote a short prompt: "An argentine woman with green eyes and blonde hair, dressed in the blue-and-white uniform of a space fleet officer." Then I used the "improve prompt" button that caused an AI to expand my prompt until it hit the character limit. The final prompt ended up as this:

"A strikingly authoritative Argentine woman in her mid 20's with blonde hair and green eyes. Adorned in a sleek blue uniform with crisp white detailing, exuding an aura of command as an officer of a space fleet. This scene is depicted in a highly detailed and realistic painting. Every intricate feature of her futuristic attire is meticulously rendered, from the shimmering fabric to the intricate insignias adorning her uniform. The image evokes a sense of power and modernity, immersing viewers in a world where space exploration and elegant uniformity merge seamlessly."

Of the four pictures generated, the one consistency I noticed was how Leonardo AI struggled with really fine details. The patches, zippers, and smaller parts of Ino's uniform look as though they've been melted or made smooth. In one portrait, Ino's eyes were also different sizes, but this error only happened once.



Deuteragonist Sota Beninato:

Sota's prompt was even shorter. I just asked for a "Singaporean man employed as an engineer on a starship." The Improve Prompt button blew up the description to:

"A tech-savvy Singaporean man in his late twenties, works as a skilled engineer on a futuristic starship. This striking portrait captures his sharp features, determined expression, and intricate cybernetic enhancements. A mix of pride and focus shines in his eyes, reflecting the glow of holographic readouts around him. The image, possibly a hyper-realistic digital painting, showcases the intricate details of his sleek silver uniform and advanced technology integrated seamlessly into his body. Its high-definition clarity emphasizes the man's expertise and dedication to his craft, making him a vivid and compelling figure in this sci-fi setting."

Every time I generated an image of Sota, he came out wearing lightweight power armor similar to what's shown in the above picture.



Tritagonist Elodie Lichtenstein:

Elodie's portrait came as a stroke of unexpected luck. Because I'm using the free version of Leonardo AI, I was restricted to using the image generator only a handful of times per day and I ran out of uses while trying to create a version of Sota that did not make him look like the T-1000 from Terminator 2. Fortunately, I had a number of rejected images left over from when I was trying to generate a portrait of Ino. One such reject became Elodie's portrait.

I've also used Leonardo to create pictures of props and vehicles that will be appearing in Unsung Paragon:



The MiG-95 Starfighter

Believe it or not, I got this image while trying to generate a picture of the battleship Marathon, which is going to be the main hero ship of the new story. The MiG-95 is going to be seen a lot in Unsung Paragons, so I decided to hold onto this picture anyway. The prompt used was: "A Stellaris-style painting of a large military starship flying through a colorful nebula." I'm pretty sure the vague nature of the prompt is the reason I got a starfighter instead of a battleship.



Hero of the Human Race

I asked Leonardo AI to give me "A prestigious medal for heroism on a regal blue and silver ribbon. The words 'Hero of the Human Race' are meticulously engraved on its surface." I forgot that AI Image generators can't really handle written words, but I still got a satisfactory result here.



All told, I generated 48 images and came away with 5 that I might use in Unsung Paragons. So far, I think creating character portraits is the best use for Leonardo AI as far as my own story goes. I might give this another attempt once The Broken Gates reaches its conclusion and I go full speed ahead on the new story.

Like @CBR JGWRR said, there's a lot of potential here, but it comes packaged with a lot of wasted time.

 

[CBR JGWRR]

Getting starfighters instead of battleships is quite common; it happens so often it was one of the reasons why AI art wasn't used as much for Life2.0. (although Life2.0 especially was limited for AI artwork for ships as the AI has no idea what Project Orion is, let alone the derivatives I was coming up with; personally I use Galactic Civilisations 4 for ship artwork)

So, following on from your post, I decided to have a go at it myself.

I should add I got utterly fed up of the waiting for a slot to have my generations fitted in and the cap on the amount of images you're allowed to generate - which I don't necessarily disagree with, given they do have substantial overheads to cover even if the marginal cost per picture is miniscule, they do have to prioritise paying customers, but it is an unfortunate outcome - so Playground AI got me to break the habit of a lifetime by being the first subscription I've ever actually taken...

I'm allowed 1000 images a day and I generally only use a fifth of that (most of my images are a lot of batches) so if you want stuff iterated, I can do so.

Anyway I copied the AI-expanded prompt and ran off a few batches of 4. The one I think is closest to Ino is this:

But...

It doesn't quite feel right.

So, I redid it using the way that works for me. The AI datasets inevitably get a lot of training on Humans, and a lot of the time you can get decent results by just entering the name; I then also added "Starfleet Captain" - the AI usually figures out that you want a space officer theme from there. Again, ran off a few batches, and the only one with greenish eyes and blonde hair was this one:


I then specified a white and blue uniform. Curiously, all the images in the batch that did not have the uniform's colour specified were blonde. Yet Playground AI seems to think Ino should be a brunette if she wears a white and blue uniform...

I must confess, after quite a lot of generations, I did resort to using the inpaint tool to give her green eyes:


I think this was the nicest uniform - the others were generally a bit OTT.

Did also have a go at the MiG-95; specifically, I put in "MiG-95 Starfighter". I was curious to see what it drew, and I found the first image: amusing enough to share:

...It reminds me of when I replicated a Spitfire - complete with spinning propeller - in GalCiv3.

 

[Macavity116]

Getting starfighters instead of battleships is quite common; it happens so often it was one of the reasons why AI art wasn't used as much for Life2.0. (although Life2.0 especially was limited for AI artwork for ships as the AI has no idea what Project Orion is, let alone the derivatives I was coming up with; personally I use Galactic Civilisations 4 for ship artwork)

Yeah, given how heavily trained the AI artist is on human subjects, I'll definitely be keeping to Stellars (and KSP to a lesser extent) for images of specific ships.

I'm allowed 1000 images a day and I generally only use a fifth of that (most of my images are a lot of batches) so if you want stuff iterated, I can do so.

One thousand per day!? That just blows my mind!

Once I get closer to launching Unsung Paragons, I may just recruit you for new iterations of the protagonist portraits. It's way to soon to say whether any of these are going to stick.

after quite a lot of generations, I did resort to using the inpaint tool

I know Leonardo AI has its own variant of this, will probably try it out later on.

 

[Chac1]

Good to see this exchange happening. I think these examples from @Macavity116 & @CBR JGWRR are exactly why various folks thought this thread might be interesting and a service to those just dipping their toes into AI art generation.

Apologies that RL has been keeping me from the forums. However, great to see this trading of information.

Perhaps, @Macavity116 you might share what you hope to know more about as someone just starting to explore this work to enhance AARs?

(I have to add, yes, there is a lot of "wasted time" while you make and sort batches. This is why I tend to save images and repurpose them too.)

P.S.: Great to get these preview images from two AARs that are developing.

 

[Macavity116]

Perhaps, @Macavity116 you might share what you hope to know more about as someone just starting to explore this work to enhance AARs?

Well, I've got three goals, an album of AI-generated images that could be used in Unsung Paragons:

1. Official-looking portraits of all six major characters (the type of headshots that might be displayed in a TV news report or in a history book)
2. "Action shots" of the main spacecraft that will play central roles in the story, mainly the Marathon, her compliment of MiG-95 strike craft, the science vessel Vivarium, and Araj Zilfallon's wolfpack.
3. Stylized pictures of certain props/McGuffins that appear in the story. (Ino's Esperanto flag, the Staff of Life, Rylan's heart*, the Time Crystal, and so on)

*Rylan Quin is a cybernetic lifeform; mostly organic with a plethora of synthetic implants. I'm thinking about using his removable/replaceable heart as a McGuffin.

After yesterday/today's experiments, the "action shots" bullet has been pretty much ruled out, but I still want to try options 1 and 3.


There are two points that I want to explore further:

• Reference images: Leonardo has a function that allows me to upload an already existing image and use it as a starting point for future iterations, but the free version is limited. (obviously) Are there any advantages to using a reference image over a written prompt? Do other AI generators have better reference image options than Leonardo?
• Uniform artstyle: I want all of the images used in my story to have a consistent visual style. (the anime style used for commissioned artwork in several of my previous AARs) I think this can be done by just adding a specific line to the written prompt, but if you know another way, do let me know. It would be appreciated.

 

[Chac1]

Not an expert but definitely somewhere beyond beginner so I will attempt to be helpful here....

• Reference images: Leonardo has a function that allows me to upload an already existing image and use it as a starting point for future iterations, but the free version is limited. (obviously) Are there any advantages to using a reference image over a written prompt? Do other AI generators have better reference image options than Leonardo?

Various platforms allow reference images now. I have used them with Bing/Copilot, Playground, and Leonardo. I do think this sharpens the AI beyond the words used. But the process isn't always perfect and in my impatience sometimes I bolt to a different platform if my initial experiments don't yield results. I usually try words only to start and then move to reference images as part of my process. Amazing what you can get with just words, and sometimes the AI provides great surprises. But if you are in a hurry, starting with a reference image cuts down on worthless batches in my experience.

Of the three platforms that I use that allow reference images, I think Playground is the best for this. I think some of the results shown by @CBR JGWRR show this.

• Uniform artstyle: I want all of the images used in my story to have a consistent visual style. (the anime style used for commissioned artwork in several of my previous AARs) I think this can be done by just adding a specific line to the written prompt, but if you know another way, do let me know. It would be appreciated.

I think you can get some of this by using a specific prompt that references a style you like, so that is why I often use a prompt that references Romantic 19th Century art or Gothic styles. However, you don't always get uniformity.

I have done experiments with RenderNet and those mostly failed. However, they have a new version out where they promise better results in their Character Lab. Have not had time to experiment with this but I will get around to it eventually.

Uniformity of style and character in AI image generation is a long-standing issue.

 

[CBR JGWRR]

Yeah, given how heavily trained the AI artist is on human subjects, I'll definitely be keeping to Stellars (and KSP to a lesser extent) for images of specific ships.

Understandable.

I did try to recreate the MiG-95 in GC4 out of interest:

Not quite right. Not that good with the software...

One thousand per day!? That just blows my mind!

Once I get closer to launching Unsung Paragons, I may just recruit you for new iterations of the protagonist portraits. It's way to soon to say whether any of these are going to stick.

Yeah; Playground AI offers several subscription packages, and that's what you get with the cheapest.

And as said, I'm happy for that - if nothing else, you can bounce off ideas for other characters. And there's no shortage of demand for Human character concept art.

I know Leonardo AI has its own variant of this, will probably try it out later on.

It is... Hit and miss. Mostly miss...

But, it can work for those times where the image is 90% passable and just needs a few tweaks. I'd say the Naomi and Buri wedding day image is fairly representative of my experiences with it and how it works.

 

[Macavity116]

I did try to recreate the MiG-95 in GC4 out of interest:

Not quite right. Not that good with the software...

It may not look or feel right, but I can definitely use this as a starting space. It's not like I'm staying in the square Stellaris painted for me anyway. In-game, all of the strike craft are referred to generically as "Strike Craft." I want the MiG-95 to be a "multirole" spacecraft. So I'll be taking this design to make some tweaks and experiments in Kerbal Space Program.

there's no shortage of demand for Human character concept art.

So I have discovered. The AI's heavy focus on creating Human characters is already starting to stump me on the Cyborg/cybernetic character Rylan, owing to the fact that he was not Human to start with. Asking the machine to give me a picture of a Cyborg Alien has made for some fun results, but no usable ones. XD

 

[CBR JGWRR]

A multirole strikecraft will definitely need more hardpoints; the AI artwork only shows missile mounts either side of the cockpit... I added what could conceivably be a nose-mount for a kinetic weapon.

The big issue it's got is the fuel tanks are inevitably going to be really small, it's probably only got a couple of kms in there for KSP's rockets.

Rylan sounds like a good candidate for starting off with a reference image, if you can find something suitable...

Lets see...

On a spectrum of say, the Terminator, Lt. Cmdr. Data and a Fallout 4 Gen 3 Synth, where does he fit?

 

[Macavity116]

Rylan sounds like a good candidate for starting off with a reference image, if you can find something suitable...

Lets see...

On a spectrum of say, the Terminator, Lt. Cmdr. Data and a Fallout 4 Gen 3 Synth, where does he fit?

LT. Cmdr. Data was definitely my starting point for Rylan.

 

[CBR JGWRR]

Ok, so where's the concept go after Data? Could you use him as a reference image?

 

[Macavity116]

Theoretically, yes. My choice of alien empires complicates the idea. Rylan's species (the Augments) are going down the Cybernetic Ascension path, so most members of that race, including Rylan, have Ritualistic Implants. (example pictured below, the Augments will be at Stage 2 of their Cybernetic Evolution during the new story)

When I ask the AI for a picture of someone modified with cybernetic implants, more of then than not I get back a picture of a full android rather than someone closer to a cyborg.

 

[CBR JGWRR]

Ok, that sounds like this is where the inpaint or equivalent feature would work better.

Another method is to make your portrait and implants separately, then combine them into one image. Playground AI has the limitation that the smallest it generates is 256x256, but you can generate it, remove the background, then rescale it. Combining them is done by dragging the implant onto the portrait then selecting them all and using the image to image feature on a high (90% or more) image strength.

 

[Macavity116]

Aha! I've seen the "remove background" button lurking about before. Glad to see there actually is a use for it.

 

[CBR JGWRR]

Aha! I've seen the "remove background" button lurking about before. Glad to see there actually is a use for it.

Yep, it does have a use.

It's also useful for exporting that particular image without the background to use in other software.

 

[Chac1]

Theoretically, yes. My choice of alien empires complicates the idea. Rylan's species (the Augments) are going down the Cybernetic Ascension path, so most members of that race, including Rylan, have Ritualistic Implants. (example pictured below, the Augments will be at Stage 2 of their Cybernetic Evolution during the new story)

Is there a reason not to use the Stellaris image as the reference image? Or do you want to range far away from any game images?

 

[Macavity116]

Honestly, using the in-game screenshot didn't cross my mind. That may be tomorrow's experiment.

 

[CBR JGWRR]

Honestly, using the in-game screenshot didn't cross my mind. That may be tomorrow's experiment.

And what did you find?