[mod] Stable Elves

[Armchair]

View File

Stable Elves

---

A few months ago I started playing around with that fancy new image generator "AI" called Stable Diffusion.  At some point I decided that creating a portrait mod for Stellaris would be a fun project that would help give me practice with writing prompts and touching up output images.  As a bonus, a lot of the horrible flaws of "AI" generated images like their atrocious hands are hard to notice when they're shrunk down to the thumbnail sized images like the ones used in a Stellaris portrait mod.

 

The goal of this mod was to create a portrait mod with a variety of different outfits where each pop type would have its own theme, but I wanted all of the pops to have the same style and look like they belonged together as part of a larger set.  I also wanted to generate a large pool of portraits to hopefully mitigate the "Birthday Paradox" problem and improve the odds that my important leaders will have unique portraits for easier identification during a game.  Originally I was just planning to use the mod for myself, but I decided to release the mod after some people expressed interest in it while I was asking for advice on how to use Stable Diffusion.  As of the version 1.0 release, there are 151 portraits split into the following categories.

 

• 21 Rulers:  They have elaborate and ornate clothing, but the outfits are completely shameless in their use of silky sheer material.  Kind of a harem girl look.
• 20 Governors:  I gave them more of a evening gown look.  Sheer fabric is still in use but as a whole they're more modest than most other pops.  I wanted the women to look a bit more mature than most of the other pops.
• 27 Military:  Admirals and Generals share this set for the 1.0 release.  The common theme when I made theme is that I quite liked the high collared dress jackets seen on the Principality of Zeon officers from the Gundam franchise.  Especially in the striking red color worn by Char.  So I tried to keep that dress jacket and experimented from there.   I eventually want to give Generals a separate theme, but I haven't decided on what yet.
• 13: Envoys:  Extremely lewd cocktail dress was the theme.  I wanted to try to evoke a feeling like the envoys are actively trying to seduce whereas the lewd outfits of the other pops were just a product of their culture.
• 51 Scientists:  The common theme is lab coats, glasses, and turtleneck sweaters.  From there there is a mix of leotards/bottomless/underboob/boob windows to try to create some variety.
• 19 generic pops:  The basic pops have a "uniform" of favoring detached sleeves and thigh highs.

 

These portraits are defined as part of a new "Stable Elf" species, so they shouldn't conflict with any other portrait mods.

 

A disclaimer on content depicted in my preview screenshots.  This mod ONLY contains the elf portraits.  The rest of the content depicted within the screenshots include:

• UI changes are part of "UI Overhaul Dynamic" from the steam workshop.  Steam workshop credits Orrie as the author.
• Namelist used in the screenshot is from "Elves of Stellaris" from the steam workshop.  Steam workshop credits Princess Stabbity, Teneb Kei, and TheGrandN as authors.
• The enslaved prepatent species is one of the species from Mugginnato's "Stellaris Vanilla Framework" mod from loverslab.
• The flag is from TheRegalSeagull's "Womb Tattoo Flags" mod from loverslab.

 

 

I know that there has been some drama regarding "AI" generated drawings in the past few months.  I don't want to discuss that topic here.  We're all here to have fun.

---

• Submitter
  Armchair
• Submitted
  03/25/2023
• Category
  Stellaris
• Requires

 

[Synthesis]

Congratulations on getting this up and working! My installation of Stellaris has been broken for...about six months, due to updates conflicting with mods, otherwise I'd give it a crack. 

[Z0mBieP00Nani]

Stable elves?...

 

You mean, elves without daddy issues??

Edited March 25, 2023 by Z0mBieP00Nani

[Armchair]

2 hours ago, Z0mBieP00Nani said:

Stable elves?...

 

You mean, elves without daddy issues??

I'm not a very imaginative man.  I needed a name for the mod.  I tried to think of something fancy and failed.   The portraits are of elves.  The images were generated by stable diffusion.  Squish the two together and you're left with "Stable Elves".  I'm not in love with the name and am not opposed to changing it at some point, but I think it's functional.

Edited March 26, 2023 by Armchair

[genericuser27]

Any chance we could get something like this for humans?? Nice job!

[Armchair]

Thanks!

 

I don't presently have any plans beyond eventually adding a unique set of "Generals" for the elves.  I simply have other projects that I want to work on IRL.

 

 

Doing something for humans would be possible, but I think I'd probably be more interested in trying to create something more exotic such as flavors of monster-girl like harpies or something.

 

I also have homework that I would want to do before undertaking any kind of expansion.   It took me a lot longer than I originally expected to put this together.  I've got a folder containing ~5.8k elf images that were rejected for one reason or another.  Outfit looked wrong, face looked dumb, was preening herself with all three hands, HANDS in general being garbage, etc.  SD is really good at drawing a pair of tits, but things can get weird really quickly once you tear your eyes away and start inspecting the rest of the image. I need to either get better at fixing blemished images or get better at improving the success rate of my outputs.  The stable diffusion community has made a lot of strides recently in training better image models that improve overall output quality and let the user fine tune outputs to match certain themes.  I've been using an outdated model since I wanted all of my outputs to look internally consistent.  I could also speed up generation times by investing in a 4090 or something, but in addition to costing $$$$ those monsters would demand that I upgrade my PSU and I'd might contemplate doing a whole PC rebuild at that point.  I technically have the money but I feel like spending it on PC upgrades would be irresponsible right now.

 

I also need to streamline the process of excising the subjects from the backgrounds, and also streamline the process of generating the DDS files.   I found a plugin for GIMP that uses an "AI" algorithm to extract the subject from the background.  When making this portrait mod, it was successful maybe 20% of the time.  Can I improve that success rate?  Does it work better with different types of backgrounds?  Should I just throw more images at the plugin and just keep the ones that play nice?  I might be able to just extract the subjects quicker by hand if I get less picky.  I was really careful with the selection and eraser tools, but I might not have needed to be so picky because the final result gets downsized to a small thumbnail for use in the game.

 

 

But back to the original question:  Right now I don't have any additional plans.  I want to spend some evenings playing games and weekends working on other projects around my house.  Once I catch up on learning about all of the new AI toys, I might revisit the idea of expanding the mod as a means of testing the different models.  I just don't have any plans to do so right now.

 

 

Here's a link to the the GIMP plugin that I was talking about.  Page 4 of the forum thread has an easy step-by-step guide for installing it.

http://gimpchat.com/viewtopic.php?f=9&t=20041

[Armchair]

I added a new upload that contains the descriptor.mod file that was missing from the original upload.

[coldheartzero]

Images look good, I'm both happy and angry at AI generated art.

 

I'm happy that people without talent are more willing to contribute, but unhappy that people with talent are being outclassed by a souless machine as far as art goes.

 

I'm also already starting to see an influx of people claiming their art is theirs when you know it's AI generated. Anyway, not the time or place for this.

 

I might use your mod in my next playthrough, they look good.

[Armchair]

For what it's worth, the very small image sizes used in the context of this mod go a LONG way towards covering up the blemishes that normally undermine the appearance AI generated output.  Even still, you'll notice a lot of re-used poses, conveniently hidden hands, and hands that are noticeably malformed even at this resolution.

 

I know that there has been a lot of drama around AI art in recent months.  My personal opinion is that ultimately the history books will eventually look back at AI art as being another tool like digital photography or photoshop.  Something that can help the artistically talented push the envelope, streamline their workflow, improve their output, and facilitate quicker experimentation with concepts and ideas to push through writer's block.

 

That said, I can understand the frustrations that some have with it.  But I can't close Pandora's box.  For better or worse, it's something that exists now.

[Alphanineteen]

Great idea, thank you for taking the time to share it with the world.

[MagnaSonic3000]

I think AI art is perfectly fine for mods for 3 reasons.

 

1. All images are uniform, instead of grabbing from like 32 different artists. And let's be honest, most anime style nude art suffers same face syndrome, so the AI being very samey is hardly an issue.
2. People with no talent can now make whatever they want, regardless if it conflicts with a potential artist or not.
3. AI art I believe comes with a public domain copyright, as it cannot be trademarked due to being a machine, and Skynet has no human rights. This means you don't have to ask anyone for permission to do anything.

Edited March 29, 2023 by MagnaSonic3000
Added more text

[Calcifire]

It's over. ?

 

But in all seriousness. A usage like this makes a lot of sense. Im curious about you time frame. If you had ambitions to do more species, how many do you think you could crank out in a month?

[Armchair]

9 hours ago, Calcifire said:

It's over. ?

 

But in all seriousness. A usage like this makes a lot of sense. Im curious about you time frame. If you had ambitions to do more species, how many do you think you could crank out in a month?

Oof, I hope that was in jest. I'm a fan of your portrait mod and how each of the characters are dripping with soul. I was a bit worried when releasing it since some artists have viewed AI generated images as attacks against their work, and I certainly don't intend for anything of the sort.

 

 

To answer your question: Honestly maybe one in a month if I stopped getting distracted and spent my free time towards recreating what I've done where each pop category has its own theme.  Although one in a month still feels ambitious to me.

 

I started work on this mod back in October, but there were periods where I didn't work on it at all. My workflow also wasn't refined since I was learning as I went and the purpose of the mod was to provide some structure and immediate tangible goals to work towards as I experimented with the different settings.

 

The basic high level steps were the following:

1. Make a prompt that can produce the desired output.
2. Generate large quantities of outputs
3. Select the best outputs
4. Touch-up defects
5. Remove the backgrounds and resize into DDS files.
6. Write the mod files that define the species.

 

 

It would usually take me a few evenings to create a prompt that produced output that I was satisfied with. It's worth mentioning that these elves are basically just ordinary anime girls with long ears, and that fantasy elves show up enough in media that the image models know what you're talking about when you ask for an elf. Basically elves were easy-mode. It would be more challenging to get decent output for depicting more exotic species. I'm not sure where the limit is though since I have seen people generate image of monster girls, so it's at least not limited to Star Trek style “humans in make up” aliens.

 

Generating the images is easy once you have the prompt since the computer can work while you sleep, but sifting through the outputs can be boring. For every one good image there can be dozens that just look dumb for one reason or another. I did eventually stop myself once I felt that I had “enough”, but the images in my mod are what I considered to be the cream from a huge pile of ~6000 images. At times I would leave my GPU generating for ~8 hours at a time. A more modern GPU than my 1080ti would be faster.

 

Excising backgrounds and resizing is standard issue stuff for many portrait mods. I was able to save some time there since I was able to specify “simple backgrounds” as part of my image prompts, so the majority of my source images have solid color or simple gradients for backgrounds. It still took me about a week and a half to create DDS files from my 151 favorites during my free time in the evening.  I was probably too picky when it came to selections/erasing the background and might be able to speed that time up.

 

So, with napkin math, I think that I'd be making great progress if I got two pop categories worth of source images done in a week. That would be three weeks to recreate the six categories that I used for my mod. However I still don't have anything unique for Generals and I also skipped Males entirely.

[MagnaSonic3000]

1 hour ago, Armchair said:

I was a bit worried when releasing it since some artists have viewed AI generated images as attacks against their work, and I certainly don't intend for anything of the sort.

Just a reminder to you, artists don't own elves. If you post elves, they have no right to say you're stealing because you can't take something that's public domain.

[Armchair]

That's not what I meant.  I wasn't talking about IP rights.  I was referring to how some artists view the sudden rise in AI art as an attack on their hobbies and livelihoods.  Some feel as though mass produced computer images insults the years that they spent mastering their craft.  It's a very emotional topic for a lot of people.  I guess I just wanted to be clear that I respect the work that artists have accomplished.

[schroecat]

This is awesome mate, well done.

 

AI tools are an incredible step in democratising cultural production, don't ever let the luddites tell you otherwise.

[MagnaSonic3000]

Honestly, the AI rise is just another industrialization era, just on a smaller scale. Did workers hate the idea of machines taking their jobs? Yeah, but it happened anyway. Any time a new form of tech comes out, it enriches some people, and sometimes destroys others. In this case, it enriches those with no artistic talent, while potentially spitting on real artists. As someone from the former camp, I can't complain. But I think it's impossible that AI won't harm someone's model in some way, and I think it's just something that's going to happen, as AI is not slowing down, in fact I'd say it's perfecting itself.

[Calcifire]

14 hours ago, Armchair said:

would be more challenging to get decent output for depicting more exotic species. I'm not sure where the limit is though since I have seen people generate image of monster girls, so it's at least not limited to Star Trek style “humans in make up” aliens.

Right. Goblins and nekos are probably good candidates. While im sure you could make some wacky stuff with ai, it might not have the necessary sex appeal. But i mean, star trek aliens are what most people want anyway. So youll be alright if you went that route. Come back in 2 years and im sure you can find a denoising algorithm that can pump out stupid sexy tapeworm girls 

 

Buddy of mine is using ai to generate card art for a tcg he's working on. I made a couple of custom art peices for him, but i couldnt possible crank out 400+ unique cards, right? Its convenient to have the computer give him some basic potion images instead of commissioning a fuck ton of artist with a fuck ton of money. Otherwise, he couldnt bring his project to life without it. But in the same way, the limitations of the tech frustrate him. He wants to create new fantasy races for his game, but the machine doesnt come close enough to the vision in his mind. Or he likes my art style, but the thing didnt train off my work, so how do? But for making a genric flaming sword card. It works just fine.

 

Edited March 31, 2023 by Calcifire

[Whoopsiedaisy]

13 hours ago, schroecat said:

AI tools are an incredible step in democratising cultural production

More like digital communism

[Armchair]

1 hour ago, Calcifire said:

Right. Goblins and nekos are probably good candidates. While im sure you could make some wacky stuff with ai, it might not have the necessary sex appeal. But i mean, star trek aliens are what most people want anyway. So youll be alright if you went that route. Come back in 2 years and im sure you can find a denoising algorithm that can pump out stupid sexy tapeworm girls 

 

Buddy of mine is using ai to generate card art for a tcg he's working on. I made a couple of custom art peices for him, but i couldnt possible crank out 400+ unique cards, right? Its convenient to have the computer give him some basic potion images instead of commissioning a fuck ton of artist with a fuck ton of money. Otherwise, he couldnt bring his project to life without it. But in the same way, the limitations of the tech frustrate him. He wants to create new fantasy races for his game, but the machine doesnt come close enough to the vision in his mind. Or he likes my art style, but the thing didnt train off my work, so how do? But for making a genric flaming sword card. It works just fine.

 

Being able to leverage AI art does strike me as something big for small developers.  Even if they just use the AI art assets as placeholders for early builds with the intent to commission art for usage in the final release.

 

 

In the past six months I know that people have made some advances in using stable diffusion that I haven't kept up with.  I know that some people are accomplishing certain things but I don't know precisely how, and frankly at the moment I'm more intent on playing stellaris and other games rather than learning how.

 

One example of something that people have been doing is they've somehow managed to create some kind of system where they can rig up wireframe skeletons that characters get drawn around so that they can produce output with a specific desired pose.

 

Another thing is that people have been making something called "loras" which can fine tune the model's outputs.  I don't know the full extent of the capabilities but I know that people have used loras to do things like mimic an artist's drawing style or create images of obscure characters from old shows.  I know that you need a pool of sample images for training the lora, but I don't yet know the precise requirements or the actual steps to follow to make one of these loras.   Loras might be helpful for generating the more exotic aliens.

 

 

I eventually want to try making a lora based off of images of the Sangvis Ferri Ringleaders from the mobile game Girls' Frontline.  I am curious if it would be helpful in making a set of images for my mod's Generals.  Basically the objective would be to have elves in dynamic poses, with monochrome outfits, and some kind of nonsensically futuristic weapons as props.  SD has had trouble with drawing realistic weapons, but honestly the weapons wielded by characters like Architect, Gager, and Dreamer are kind of pants-on-head in the first place.

[ellylanster]

Sorry to ask, especially if it's been asked before.. but do you happen to have the tags you used to make this? I've been thinking of making something similar but I can't seem to get the images fully within frame.

[Armchair]

The main thing is to dedicate part of your prompt to describing how the shot should be framed.  For example, specify if you want a full body shot.  My prompts called for "cowboy shots" which is basically from the mid thigh up.  It's still pretty common for the tops of heads to be cut off, but that's pretty easy to fix if you read up on using img-2-img inpainting.

 

Below, within the spoiler tags, are the positive and negative prompts for one of the military officers.  Originally I copied someone else's prompt and just edited from there.  Sometimes I feel like a cargo cultist who's doing things for no good reason just because it seems to produce results.  For example, the original prompt that I copied contained a token for "chibi" which was odd since I didn't want chibis and the outputs mostly weren't chibis.  But I felt like the outputs were better if I kept "chibi" in there.  It might be entirely placebo but I thought that the presence of chibi made the characters look softer/more hugable.

 

I also don't really know if it's better to use commas as delimiters or just separate tokens by spaces.  I used underscores when trying to strengthen the bond between tokens, like for example a red_jacket.  But I didn't have full success and underscores may very well simply not work that way.  It's definitely not a guarantee that joining tokens with an underscore would "link" them like that but subjectively I felt like it stacked the deck towards your the output working like that.

 

It's also a long prompt, which I understand to be poor form.  My understanding is that in general terms your first 75 tokens in the prompt follow a descending trend of weight where the first tokens in the prompt are more influential than the last one.  However if you overflow into 76 tokens and beyond to 150, then the 2nd batch is basically split off by itself and follows a new descending ranking of weight.  Then the two sets are kind of ambiguously merged.  That was the layman description of how long prompts worked, and that behavior could make long prompts a bit unpredictable if you start re-arranging the prompt and tokens near that dividing line get moved between the two sets between generations.

 

In general I recommend looking at images on danbooru and then seeing what kind of tags were used to describe the images.  Then use those tags as tokens in your prompts.

 

I also recommend playing around with X/Y plots to experiment with comparing the same prompt with different variations of other settings like sampling method, CFG scale, denoising, or sampling steps to see if anything "looks right" to you.

 

The upscaling options tended to introduce more detail as it filled in the gaps between pixels, but that obviously takes longer to run.  To save time I tried to make low resolution rough drafts with different prompts before then experimenting with upscaling the results from prompts that I liked.   I also generally would test each prompt with a small batch rather than an individual generation just in case rotten outputs were the result of a bad seed rather than the result a bad prompt.  

 

Positive/negative prompts within the spoiler tags because they're kind of a mess to look at.

positive prompt:

Spoiler

masterpiece, highest quality, cowboy_shot  high_angle (skimpy:1.04), (short:1.15) and (toned), (chibi:1.04), (dark elf:1.08) slut with (medium breasts:1.1), ornate_intricate_red_crimson_high_collar_military_dress_uniform_jacket (impossible_clothes:1.2) see_through_transparent_dark_slate_bodysuit_plugsuit covered_nipples_poking_through_clothing lowleg_microskirt (pussy_peek:1.1) (womb_tattoo:1.1) (no_panties:1.1)  thigh_highs , seductive_smile  assertive female, eyes up, lips,   simple background  gold_trim

negative prompt:

Spoiler

multiple girls, lowres, text, error, missing fingers, (bad anatomy), extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, extra fingers, mutated hands, (multiple penises), (disembodied), (poorly drawn hands), (poorly drawn face), (mutation), (deformed breasts), (ugly), blurry, (bad proportions), (extra limbs),  blowjob, deepthroat, animal ears, extra ears, panties

 

Edited April 1, 2023 by Armchair

[Særtr]

Not sure what the issue is but I seem to only see StableElfRuler00 for everything.

[Armchair]

Ruler00 is the default portrait, but you should be seeing the rest of the portraits too.  Their definitions are in the gfx\portraits\portraits\00_StableElf_portraits.txt file.  I suppose maybe you might just see the default picture if you accidentally edited and broke the rest of the file.  The SHA-1 hash of that file is ef8f7003f4d203e11aff1b51a3fa0a94f89d6815 if you want to check to see if something changed.  Or maybe you could delete the whole mod and re-extract everything.
 

I suppose maybe another mod might be causing some conflict, but I'm not sure what would do that since my species was defined as a new unique archetype.  There's always re-creating your mod list one by one.  It's a bit of a drag but sometimes that's what it takes to find the problem.

[Særtr]

Got it, turns out it was another mod, not a conflict but a feature of the mod that was poorly communicated.