AI Generative Art

[Braddack]

my newerst Lvl 50... Lightning Cloud

 

[Palehood]

For those using KREA, is there a way to "edit" or "tell" the AI to make the Enhanced face look to be of a certain race, cause there was once an asian looking character turned out looking white after I tried Enhancing the character.

[Marshal_General]

40 minutes ago, Palehood said:

For those using KREA, is there a way to "edit" or "tell" the AI to make the Enhanced face look to be of a certain race, cause there was once an asian looking character turned out looking white after I tried Enhancing the character.

I believe you can click on the right side to expand the options and add prompts to it along with some sliders to play with.

[Player2]

33 minutes ago, Marshal_General said:

I believe you can click on the right side to expand the options and add prompts to it along with some sliders to play with.

Yes, exactly.  There's a button that says "Settings" below where your uploaded image is.  When you click it, there's a box for Prompt which usually auto-generates to describe the image you uploaded; If it doesn't, there is a little Auto-Generate button in the bottom left corner of that box. In this box, I usually specify things like "Asian Woman" or "Robot Man" as appropriate.  Don't assume it's always correct, either.  I've noticed the auto-prompt give me partially incorrect descriptions like calling a blonde a redhead or giving me weird descriptions of things in the background that just aren't there.  Fix it here.

 

After that, there's a drop-down menu (Default, Flat Sharp, Strong, Reinterpretation, Oil Painting, Digital Art), followed by some sliders (AI Strength, Resemblance, Clarity), a Color Correction button, and another prompt box called Negative Prompt where you put in things that you don't want to see.  This one is going to be hit or miss depending on how much you mess with your options, but it's not a miracle magic deletion button that will make anything you don't want go away... I wish, though.

[Trike]

3 hours ago, Palehood said:

For those using KREA, is there a way to "edit" or "tell" the AI to make the Enhanced face look to be of a certain race, cause there was once an asian looking character turned out looking white after I tried Enhancing the character.

As Player2 says, you can specify things in the prompt box under Settings.

 

I added “Korean woman in her 20s wearing a traditional Korean outfit” to the prompt for the enhanced version of Mirror Spirit:

 

[Trike]

1 hour ago, Player2 said:

I've noticed the auto-prompt give me partially incorrect descriptions like calling a blonde a redhead or giving me weird descriptions of things in the background that just aren't there. 

It is a bit strange how it sees things. In the Back Alley Brawler one it said he had his arms crossed. No idea how it came to that conclusion.

[Cyri]

 

I cannot seem to do backgrounds well as Krimson can.  The above is the result of literal hours of loading files into Krea and continually adjusting the prompts to achieve better results.  My opinions on it are...mixed. 

Edited April 25 by Cyri

[Cyri]

 

 

At this point, I believe I have gone as far as I can today until I can think of new ways to describe what I want produced.  These preceded those from above.  Not sure if any of them are good, or which are better.  I mean, they're not terrible, but I need to work on this if I am going to get the desired results.

[Yodafan07]

[Cyri]

 

Strangely, some work well, and some do not.  This one has pleased me. 

[Krimson]

Prompts are your friends. Don't trust Krea to auto-generate them.

[Trike]

3 hours ago, Cyri said:

At this point, I believe I have gone as far as I can today until I can think of new ways to describe what I want produced.  These preceded those from above.  Not sure if any of them are good, or which are better.  I mean, they're not terrible, but I need to work on this if I am going to get the desired results.

 

I've found that with all of these AI programs sometimes the order in which you phrase things has an effect, but I'm not entirely sure what the difference is. It's very much a black box.

 

For instance, I did a bunch of anthropomorphic characters and I wanted them all to have a variation of the game's paw print inside a circle as their common feature. If I put that first, I got paw prints all over the picture, but if I put it last then 9 out of 10 times it added them where I wanted. Sometimes it surprised me, like when it put the paw print on a dog's collar and on a polar bear's belt. 

 

After many iterations and squeezing my brain for thesaurus words, I found that "emblem" was the best word to describe what I was looking for. "They have a chest emblem that is a stylized paw print inside a circle." But it took me dozens of attempts to get to that solution.

 

I didn't keep the ones that were far misses, but I kept a few examples.

 

I asked for an anthropomorphic wolf wearing green tights with the paw print. This was the result. Why it rendered a female with long hair I have no idea, probably because the furry community does this, but it was getting close to what I wanted.

 

Same with the caracal, except this time I used the word "emblem".

 

So close!

 

So then I tried moving that specific ask -- "a chest emblem that is a stylized paw print inside a circle" -- to different places in the prompt. Once I put it in last place I got consistent placement with my "wolf wearing green". Except it kept giving me lady wolves with hair. (I also removed the "wearing brown boots" ask because it looked dumb.)

 

 

So I specifically asked for a male wolf, and that did the trick.

 

After that I simply swapped out different animals and colors. Instead of "wolf" and "green" I did "chimpanzee" and "yellow", or "lion" and "maroon", etc.

 

 

This was dozens of attempts over a couple weeks, and it was frustrating... until I broke through with the right word placement. Then it was just a matter of plugging in different animals to see what I could get. The program absolutely refuses to make a bull, again probably because of swingers, but substituting buffalo and bison worked. Then I started messing around with camera angles and I decided to try to do a version of the classic Justice League cover where they're all looking up at the viewer.

 

And that's how I ended up with a massive desktop wallpaper-sized version of a bunch of animal characters. I rendered them individually, then took them into a free program called paint.net and pasted them together. It was a lot of work, and i actually forgot to include a couple. Again, check out the bear, panda and Akita on the left side. The program can sometimes surprise you.

 

(Click the pic to see the full size.)

 

Edited April 26 by Trike
clarified a thing

[Krimson]

14 minutes ago, Trike said:

 

I've found that with all of these AI programs sometimes the order in which you phrase things has an effect, but I'm not entirely sure what the difference is. It's very much a black box.

 

For instance, I did a bunch of anthropomorphic characters and I wanted them all to have a variation of the game's paw print inside a circle as their common feature. If I put that first, I got paw prints all over the picture, but if I put it last then 9 out of 10 times it added them where I wanted. Sometimes it surprised me, like when it put the paw print on a dog's collar and on a polar bear's belt. 

 

After many iterations and squeezing my brain for thesaurus words, I found that "emblem" was the best word to describe what I was looking for. "They have a chest emblem that is a stylized paw print inside a circle." But it took me dozens of attempts to get to that solution.

 

I didn't keep the ones that were far misses, but I kept a few examples.

 

I asked for an anthropomorphic wolf wearing green tights with the paw print. This was the result. Why it rendered a female with long hair I have no idea, probably because the furry community does this, but it was getting close to what I wanted.

 

Same with the caracal, except this time I used the word "emblem".

 

So close!

 

So then I tried moving that specific ask -- "a chest emblem that is a stylized paw print inside a circle" -- to different places in the prompt. Once I put it in last place I got consistent placement with my "wolf wearing green". Except it kept giving me lady wolves with hair. (I also removed the "wearing brown boots" ask because it looked dumb.)

 

 

So I specifically asked for a male wolf, and that did the trick.

 

After that I simply swapped out different animals and colors. Instead of "wolf" and "green" I did "chimpanzee" and "yellow", or "lion" and "maroon", etc.

 

 

This was dozens of attempts over a couple weeks, and it was frustrating... until I broke through with the right word placement. Then it was just a matter of plugging in different animals to see what I could get. The program absolutely refuses to make a bull, again probably because of swingers, but substituting buffalo and bison worked. Then I started messing around with camera angles and I decided to try to do a version of the classic Justice League cover where they're all looking up at the viewer.

 

And that's how I ended up with a massive desktop wallpaper-sized version of a bunch of animal characters. I rendered them individually, then took them into a free program called paint.net and pasted them together. It was a lot of work, and i actually forgot to include a couple. Again, check out the bear, panda and Akita on the left side. The program can sometimes surprise you.

 

(Click the pic to see the full size.)

 

Somewhat related. I made these yesterday.

[Player2]

The MindBrain is online...

 

 

NecroAngel

 

 

Big Bad Baby Eater is craving a little snack...

 

 

Floodstorm is all wet

 

 

Crafter has something for you.  It's pain.

 

 

Viridian Storm... because Green Lightning sounds both dull and derivative.

 

 

Mister Mojo will have you rising with his gravity tricks.

 

 

We do not talk about the Big Guns.

 

[Cyri]

59 minutes ago, Krimson said:

Prompts are your friends. Don't trust Krea to auto-generate them.

 

Thank you  🙂.  I have played with them quite a lot.  What I meant was that I seem to have exhausted my imagination for prompts for a bit.  After several hours (and I had fun, mind you XD), nothing I did seemed to offer any kind of meaningful improvement. 

 

What I HAVE noticed is that the style of art one elects to choose seems to alter the results significantly (as it should) but I was not aware of the style thing at first.  I went with default.  Then I noticed it and made changes.

 

I would ask you whether you have noticed the AI Strength and Resemblance getting better or worse as they're increased or decreased?  That is, put more simply, where do you feel you get the best results.  You and P2 seem to generate good results.  I have followed your character generation with some interest for a bit now and I do wonder how you do it.  I am aware of the prompts, but what else do you do? 

Thank you for your time  😊.

[Techwright]

On 4/24/2024 at 1:39 PM, Palehood said:

I posted this just for the lols before but this time I think I'm really going to really create a yet unnamed character for play this time.(Why did Krea give him human

hands(?!)

 

 

I, for one, would dub him "Steak Knife".  😉

 

 

On 4/24/2024 at 9:03 PM, Trike said:

Mighty Man
 

 

Bright Tiger and Bamboom

 

 

On 4/24/2024 at 9:11 PM, Trike said:

Major Mighty and Mighty Lad, 1950
 

 

 

Did everyone give up on including description details for training and comparison?  Trike, I'd very much like to know what the text was that rendered these art styles.  They're very much in keeping with what I'd like to achieve on some projects.

[Trike]

1 hour ago, Techwright said:

Did everyone give up on including description details for training and comparison?  Trike, I'd very much like to know what the text was that rendered these art styles.  They're very much in keeping with what I'd like to achieve on some projects.

I’ve just become lazy. 😛

 

I’ve been doing so many iterations that I haven’t been writing them down. These are easy, though.

 

The top two are “Japanese brush painting, in the style of Yoji Shinkawa, (fill in character description here)”.

 

The bottom one is “oil painting, Norman Rockwell style, (character description)”.

 

Unlike Krea or DeviantArt’s Dream Up, Bing AI still works after you’ve used up your 15 credits. But I’ve never had to wait more than a minute when out of credits, so I just keep iterating while doing other things. Tuesday I did them in between Penelope Yin TF missions for Tanker Tuesday, last night it was while we watched TV, tonight it was between Halo matches.

 

Tonight I kept trying to get it to make me a picture of a wingless Chinese dragon fighting a flying European type of dragon, but no matter what prompts I gave it the dang thing kept responding with the same incorrect image. So I burned through my 15 credits and then just keep trying prompt variations.

 

 

[Trike]

On the topic of “Japanese brush painting, in the style of Yoji Shinkawa”, for some of them I asked for a specific item in a particular color, and the results were gorgeous. They weren’t CoH-related, so I didn’t post them, but here are a few:

 

The above prompt with “medieval knight with a purple cape”

 

 

+ “squire wearing orange cloak carrying a bunch of swords like a golf caddy”

 

 

+ “goblin wearing a red cap welding large knife”

 

 

+ “gryphon with golden feathers and a human for scale”

 

 

+ “wizard wearing a pointy blue hat”

 

 

…you get the idea. But I mean, check it out: I did hundreds of these! That’s why I haven’t been writing down my prompts. XD

 

 

[Krimson]

A lot of it is RNG, with both SD 1.5 and SDXL. Also, sometimes Loras don't play nicely with one another. So when I generate images, I expect to discard most of them. Though when I find a combination that works, then I can do a batch.

[Trike]

Speaking of Norman Rockwell style, here’s a sampling of some I had Bing make using that prompt:

[Krimson]

Blue's creator, Crey biotech scientist Dr. Aoi Aozora, after having a conversation with a member of the Watch, and watching him leave shaken.

[Force Redux]

On 4/25/2024 at 6:42 PM, Krimson said:

Prompts are your friends. Don't trust Krea to auto-generate them.

 

this one came out very natural, I really like it 

[Player2]

Time for some Krea Giant Monsters...

 

 

 

 

 

 

 

 

[Krimson]

15 hours ago, Force Redux said:

 

this one came out very natural, I really like it 

I think that was done with an SDXL model, my Blue Idoru v3 LoRa, and then upscaled on Krea.

[Techwright]

On 4/26/2024 at 4:32 AM, Trike said:

Speaking of Norman Rockwell style, here’s a sampling of some I had Bing make using that prompt:

 

Bing engineers must have made some adjustments to the A.I.  I've not noticed any extra or distorted figures in your pictures of late.