Fundamentals of AI Art Generation

[Chac1]

Some may have noticed a deep dive into AI art posted in the SolAARium recently.

After some debate, the thought was some of this discussion should move out of the SolAARium and into the general forum. But as @Lord Durham noted in that discussion, this thread shouldn’t be just for posting pretty pictures. The concept here is to post some of your art to help instruct others about prompting and the use of different AI image generators. Perhaps there should be an interchange of ideas here.

So let me kick this off by further discussing some of the art in that SolAARium post and expand a bit on what was discussed there. (The ethical part of that discussion is still going on there, and it has expanded to cover AI not only for making images, but also music and writing AARs too.)

Getting AI to work for you in your AAR boils down to using words as prompts to get the AI to make the images. In that post, I glossed over how I had created one piece of art for the latest chapter in my AAR.

For those who are interested, here’s more of a general step-by-step approach. My platform of choice is Playground. So I go there and pick the Stable Diffusion XL image generator. I can choose several different image generators but this is the one I like best. Then I punch in this for the prompt: “A large African man, six and a half feet tall, who is a guard armed with a scimitar dressed in a long flowing white robe painted in the Romantic style of Danish artist J. L. Lund.”

I use the filter style: “Masterpiece” which adds a variety of additional items to the prompt.

I also use a negative prompt. For those who don’t know, a negative prompt supposedly stops items from appearing. But it doesn’t always work. Here’s the negative prompt I used: “ugly, deformed, noisy, blurry, distorted, out of focus, bad anatomy, extra limbs, poorly drawn face, poorly drawn hands, missing fingers, nudity, nude.”

Here’s one of the better results of the four images produced:

Here’s a link to the work so you can see all the settings.

Not exactly what I asked for in the prompt. What’s that giant spear all about? Not to mention the disk hanging from his belt.

Five rounds later, I still don’t have what I want.

You can see from this link to my prompts, I have adjusted the main prompt and changed the filter to “royalistic” but it isn’t improving. What you see below is one of the best of the more than 20 images produced.

I decide to move onward to a different platform: Leonardo AI.

On Leonardo, I pick the Leonardo Diffusion image generator (there are various you may choose) and I pick the Digital Painting filter. However the links on Leonardo only take you to my finished image, not the prompts, so I will include a screenshot that shows you the information after the finished image. You can see I have expanded the prompt I’m using beyond what I used on Playground and greatly expanded the negative prompt too. Here is one of the first results:

Still not what I want. You’ll see after another round with Leonardo, I adjust the main prompt and the negative prompt.

The issue is the curved sword. Neither of these platforms wants to produce the sword.

Here’s what I tried on the last and final round, with the instructions following.

This is when I turned to Google’s experimental ImageFX. The problem with ImageFX is that if you don’t store the URLs of your images you can’t see the prompts. There’s no account with a bank of images like the other services offer. Sorry to say, I did not save the URLs in this experiment so you can’t see this with exactitude. However, I used a prompt similar to this:

“A large African man, six and a half feet tall, who is a guard armed with a curved half-moon-style sword, a scimitar, dressed in a long flowing white robe inside a beautiful Arab palace painted in the Romantic style of a Danish artist.”

This time, I get three possible images to use:

I decide I like the last one best and use it for the post.

However, in replicating this experiment in the preparation for this discussion, I find that ImageFX has improved its quality of what this prompt produces and two of the images are similar to this first round. These I can share with you along with their URLs so you can see how this works a bit more. Here are the two best, starting with this one:

Now, the second and likely the best of the lot:

For those who are new to AI art, you can see, this relates to playing with the words to get the right combination, along with finding filters and platforms that might work best for you.

Perhaps others will post some of their art and share their process so folks here in AARland can learn from our experiments.

 

[Chac1]

This was also a problem with my first AAR, since I wanted to make a drawing for each chapter. Made it much harder to retain a regular update schedule and, as far as I can tell, did not add significantly to the whole.

About halfway through The Patriots, the AAR about Argentina in HoI4 from @Captured Joe . I see lots of interesting images there but most appear to be from the game or in the public domain. Worth checking out, one way or another.

***​

Wanted to add something here about image generation after getting a question from @Rensslaer about this. I did give an initial answer but I thought this might be worth expanding on for those with no experience with AI image generation. Those who have some experience with image generation will already know about these basics.

I see the question as two-fold. First, @Rensslaer is correct that if you use a reference image that there are no guarantees you can replicate the look of a character or subject. Some AI generation platforms make that clear with various warning caveats. However, some platforms are advertising they have solved that issue. I have not had time to investigate those claims. Perhaps that will be a post from me or others later.

But the second issue is that usually, you can set an image generator to give you multiple examples to use, often from two to four images. These images are sometimes astonishingly alike so they appear to be of the same subject. That won't be the case in my examples below. However, just know if you haven't tried one of these before, you will get multiple images and depending upon your time and the platform, four images can quickly become 20.

For now, we are just going to use three from a recent experiment I tried on Leonardo AI. Not sure if or when I will use these images of a dog in my AAR but they will serve as a good example. These are all from the same prompt:

Here's the prompt: Inside a wooden longhouse in the 8th Century, a full-grown black and white Danish farm dog growls, drawn in the style of Hal Foster like the Prince Valiant comics.

I used a long negative prompt too, that I won't share.

These are relatively close to what I want. The one on the far right is closer to the look I wanted but I don't like the background. I like the background on the one on the far left best. So I use Leonardo's digital editor to try to adjust the background: eliminating the window and the grass. (How did that get inside a house?)

After editing and using Leonardo's image "upscaler" here are the results:

Far from perfect but acceptable.

Hoping now this gives @Rensslaer the complete answer I could not provide earlier and that others who are just beginning to dabble with AI image generation start to see what these image generators can do and what their output is really like considering the prompt provided.

 

[Chac1]

A Further Analysis of AI Art Platforms

This is an early draft of an AI art project that will likely not be used anytime soon. It is one of my early experiments carried out with the Bing image generator.

Following up on the recent request from @Lord Durham , this is another attempt to review some of the platforms for creating AI art. AI art has proven to be a very dynamic field with many changes just within the past six months. One of those changes was the recent demise of my favorite platform.

That said, please realize this is a very subjective review and it is a bit of a sequel. These thoughts actually started with this post from the SolAARium. The first post in this thread also has some aspects of this review.

As pointed out in the recent note from @CBR JGWRR , my old favorite and stand-by platform Playground has changed into something else and dismantled its best AI canvas features for creating art. I’ll be downloading my cache of visuals from that site before February so I can retain many of the visuals I’ve created there.

So what’s left?

Some may remember I started with Bing’s image generator which has now become Microsoft’s Copilot AI. I liked a lot of the quality of these platforms but got angry when images I had stored on the site were erased. Now, they clearly post that images will disappear after 90 days. However, in returning to the site after many months, I see many of my older images were magically restored somehow.

This doesn’t mean I’m returning full-time to Copilot. I will also be working to make sure I have all of my images properly saved from this site as they have already proved they are inconsistent with security for images. They also downgraded the abilities of their image editor Designer which was a big previous draw. Finally, as CBR JGWRR notes, some sites don’t let you own the images. It is clear you are allowed to use the images you generate but Microsoft owns them.

Despite those drawbacks, the site is free and the quality is sometimes tremendous, although also inconsistent. This was the first AI site I used and I find it easy for newcomers. So it may be a place to start.

Leonardo.AI was becoming a fallback location for image creation for me and that site has gone through some changes too, adding a variety of new features. CBR JGWRR uses that one because with the premium services you can own the images. One drawback I see to using the free service now is they have eliminated easy ways to delete old images! So if I want to reorganize and get rid of images I see as duds while reducing organizational clutter, I can’t do it without using a paid version.

Leonardo uses a token system for creation. The more complex the image, the more tokens. Even with the free service, you should be able to generate a dozen or so images during a session, as you get free tokens daily. But now the paid service allows you to store the tokens, depending upon how much you pay per month. The free version doesn’t allow you to bank unused tokens. AI art can take many generations and many images so a dozen is really just getting started.

The $10 per month (apologies, prices are only in U.S. dollars as that is what the site shows me) service gives you almost twice the daily tokens and a host of special services, including ways to organize images. The $24 per month service gives you six times the number of free tokens, a larger rollover token bank and other improved features. The $48 per month service gives you 15 times the number of free tokens, a larger token bank and many more high-end features. Be aware these may be holiday prices.

Not being a paid customer, can’t say if this is worth it, but CBR JGWRR seems to have had a good experience.

As I have written before, I had started to dabble with ImageFX, an experimental service from Google. They have now rebranded that as part of the AI Test Kitchen. This is a free service and don’t count on the system storing your images. Google says you own the image and the best way to stay connected to them is to copy their URL, as Google assigns an address to the creations.

Because I’ve been playing around with the backstory of Mjölnir in my AAR, as a test I used this prompt in ImageFX: "A 35-year-old beardless blonde Danish king in the 8th Century wielding Mjölnir, the sacred warhammer. Painting."

​

You’ll see the results look a lot like Thor. I wonder what Marvel would think about that hammer? Clearly, the platform did not follow my instruction for “beardless.” I’ve learned if you ask for Danes, or Norse, or Vikings, you had best ask for beardless if the character has no beard.

ImageFX has a rudimentary image editor and using it, I quickly got rid of the beard. Here’s the result:

The quality of these images has improved since I first played with this platform. Google says it is coming out with a Gemini Advanced image generator which you might have to pay $20 per month to use. I have not used Gemini to create images. My advice for now is to stick with ImageFX, which is free. (All Gemini image generators are currently free too.) The interface for image storage may be a bit awkward but you should be storing the images on your hard drive for safety anyway.

Admittedly, I was a bit lazy when I started making AI art and just used the services to store the dozens of images created. ImageFX is relatively easy to use with word prompts and the results are improving.

I have also written about RenderNet in the past. RenderNet boasts that it can consistently give you images based upon a photo of a character, putting that character in new settings and poses. In my experiments with the free service that doesn’t happen, but I was willing to try again with their improved platform. So King Þorolfr found his image soon uploaded to RenderNet.

​

This was the prompt: "^Thoroflr^ a Danish king in the 8th Century, holds Mjölnir, the sacred warhammer, painted as if in a Romantic painting from the 19th Century."

This was the best result:

Not sure about that hammer, and then there’s that nagging beard to deal with again. But that’s on me. I forgot to ask for "beardless." I had thought that uploading the picture of what I wanted would guarantee me a cleanly shaven king.

Here’s another try from my second pass, but the king still has stubble and frankly looks less like the original. And that hammer looks like it has merged with a mace.

The free version only allowed me about ten images for this experiment. I did not find RenderNet lived up to its promise of putting my character into other settings. However, your experience may vary and I do wonder what would happen with a paid plan. RenderNet has plans starting at $7 per month and ranging up to $79 per month. Not willing to pay to experiment yet when I can find other options.

There are many other services out there and I have tried three or four others. They are not worth mentioning here as I found them too complex or limiting in their free versions. However, perhaps others have had better experiences and would want to share.

Good luck and have some patience if you are venturing into AI art creation for your AAR.

 

[Chac1]

Reference Images, an Update

(As you might be able to tell from the watermark in the corner, this is from Bing. This was an early experiment using a reference image.)

Lately, a discussion elsewhere has prompted this update about using reference images. Some may recall we covered a question about such images earlier in this thread from @Rensslaer .

For the early beginner, let's define: a reference image is another image that can help the AI image generator. Image generators that allow reference images will have an interface to allow you to upload an image. This is not to copy the image exactly but to use many of its properties. Ethically, you should be using images that are in the public domain or images where you own the copyright. Or you could also use other AI images. We'll get to that in a bit.

Let's go back to the opening image in this post. I might use a cropped version of this sometime in the future, but so far haven't found that it works for me. That one is actually from 2023 using the following reference image:

That is actually a replica of the Viking Domesday Stone from Lindisfarne and it is in the public domain. The experiment used this image with additional word prompts. Our starting image was one of the better results.

One of the reasons to return to this topic is the very good AI images that @Bullfilter is creating for his Hearts of Iron 4 AAR Poles Apart. Some of his latest images include Soviet prisoners of war and Polish troops in his alt-timeline of 1947. He discusses some of his process here. That process includes something he calls a "double iteration:" using a manual edit or other electronic method to paste an image into a historic photo to alter it then using that image as a reference image in a further AI generation to as he says "smooth out the edges." You may want to check out the last few chapters of his AAR to see the examples.

This also brings up the idea of having AI art fix issues by making new generations or by reusing some of the discards from earlier generations and experiments. As I have been away from this process for more than a few months, @jak7139 reminded me that regenerating an image is always a way to fix a less than perfect piece of AI art.

For example, the pieces below:

The image on the left was used as the main chapter image for the latest chapter of Lost Seasons of the Danes. However, Jak rightly called out the horse in the background. Is that a horse or a steer? Does it only have three legs? (Look closely and you will see a white fourth leg but such wonky background images are unneeded distractions.) I had wanted the image to match a paragraph in the story that describes a springtime ride. The solution for making a better picture was using the former image as the reference image, plus better cropping of the image. The resulting image on the right is now the image used in the chapter instead of the original. These were all created with Microsoft CoPilot.

Some might ask, why not just use Photoshop or another program to get rid of the offensive horse. That too is a solution. Some of us are not so adept with Photoshop though.

Along those lines, some good news about Microsoft's Designer. Designer is a very good online editing program and perfect for pairing with images you create with Bing or Microsoft CoPilot. I was moving away from using Designer and Microsoft altogether because they had discontinued the very powerful image editor and eraser that had been included in Designer. The good news is that image editing tool has been restored. I'll be using that soon in forthcoming chapters. If you are looking for a great online image editing tool that is easy to use, that is one of the best, now that it has been restored.

Good luck with your image experiments and with using reference images.

 

[Chac1]

A few updates for folks who follow this thread....

First, you can see some of the art that @kakom has created in the third installment of the mega-campaign tracking Taurica: Taurica Part III: Age of Thought and Discovery.

Also, there was this from @Sanvone ....

Anyone has any experience with recent (2 days old) "Consistent Characters" feature introduced in many AI platforms? Cause I've seen it today and remembered that some people were complaining about not being able to replicate their characters between generations. From the tutorial it seems pretty straightforward.

So perhaps other platforms besides Leonardo.AI have tutorials posted on "consistent characters." Yes, I found the Leonardo tutorial easy to understand and I wanted to use it, but discovered the feature is not available until an unspecified later date. Perhaps if that works the way it should I will then decide to pay for a platform.

But as an experiment, I decided to generate some art with Leonardo hoping it would get closer to providing a "consistent character."

To those who might be following my new spin-off AAR, Lost Myths, you may be familiar with this avatar for Chieftess Alvör:

​

This was an image created strictly through word prompt in the now defunct parts of Playground AI. This was my reference image in my experiments with Leonardo. The fact that this image came from a different platform is immediately going to cause issues with its replication as Leonardo warns in its "consistent character" tutorial. Still this was an experiment. Consistent characters works better with photorealistic subjects. For that reason I asked for Leonardo to create output that was photorealistic. The best of the four outputs I used in the latest post even though it was far from perfect.

​

In the post, I referenced this as the Disneyfied version of the chieftess. I had tried to place her at a feast and the AI did do that, so that was a welcome outcome. However, the chieftess seemed to have her hard edges sanded away and that wasn't for the best, in my opinion. That post also includes a photorealistic profile of the chieftess that I used with some prudent cropping. However, none were truly "character consistent."

Although it was too late to include in my posting for this week, last night while creating some other pieces, I returned to this output and tried a much more detailed verbal description. Judge what you think about the below image, but the lesson for me was to try building a more detailed verbal accompaniment to the reference image, and you may have better results. However, this is still not character consistent.

​

If folks know of AI platforms offering true "consistent character" features, please share if you will and that will certainly help some of us. I will definitely let folks know when Leonardo opens up that feature and I will also test it as I've done with other platforms.