#GemsofWarCharacterIdeas 
4 Likes
Ah, interesting. I didn’t consider using “taur” on its own. Funny that the AI didn’t like centaur but does like taur.
Thanks for that. Next time I try to make one of the centaur troops, I’ll remember that.
Shows my point, though, that sometimes the AI doesn’t understand words we think are clear.
4 Likes
Absolutely correct. It’s like training a cat – which is to say, it’s really the cat who trains you. (And when I say I’ve been throwing a lot of prompts at a generator, this is what I mean by that.)
1 Like
It even likes random taur prompts, like:
Pineapple -
Sherlock -
Gnome, surrounded by treasure -
Battlecrasher -
Emperor -
And unfair -
4 Likes
Saw you shared this:
I found this issue a lot with any prompt that features multiple figures.
The generator seems to have the ability to define terms and use them in a prompt, but I’m struggling to get this to work as a means of avoiding this.
For example, I thought this would work:
M1 = a tall taur with large horns, male, with brown fur and brown armour
F1 = A short gnome with red clothes, female, red hair
DESCRIPTION = A [M1] standing near a [F1].
So, this should produce a tall centaur standing next to a short gnome, as described, and because the prompts are separated they wouldn’t mix together, but it didn’t work that way. It can still swap the colours, clothing, sizes, features, etc, giving for example a short centaur with red fur and brown clothing next to a large gnome with brown hair and armour and large horns.
If there was a way to separate the components of a prompt like this, it would probably make it easier to assign details to one figure and not affect the other. Not sure how to do it, though.
Perchance’s generator allows you to use the site’s own markup to a degree, but the actual results still depend on how the AI reads the final prompt internally, which (due to the nature of “natural language” parsing) we can’t really control. In creating the final image in that example I also attempted to specify certain elements as part of the “negative prompt” (e.g. prompt “female sergal, male protogen” with negative “male sergal, female protogen”) but it’s unclear whether it parses negative phrases the same as in the normal prompt, and it didn’t actually improve the character designs any. (As a separate experiment, I gave it various prompts of a soda can sitting on a balcony overlooking a valley between mountains, with a negative prompt of “mountain” and/or “dew”, but the results were … inconclusive at best)
The ability to label certain regions of the composition and restrict certain portions of the prompt to them would be VERY powerful indeed (and for all we know might already be in the works for new versions). For comparison, I’ve heard that DALL-E can reliably generate specific text when correctly specified in the prompt (otherwise, it may produce accurate lettering but not readable words).
PS - Perchance’s generator using a maximum image dimension of 768px suggests it may be running on Stable Diffusion 2.0 (prior versions were limited to 512px).
1 Like
I love that, though.
1 Like
Astral Spirit. Glad I got that ghostly appearance somewhat right.
6 Likes
That MAY be Copilot-specific, since Copilot appears to be first and foremost a search/chat AI, so you need to distinguish image prompts from other queries.
Having modded one of Perchance’s generators, I can now say the following for certain:
How it works
Their content filter is actually part of their “image gallery” plug-in which the generators incorporate. Individual generators can actually control whether the filter is enabled or disabled by default (per generator), and apparently this is merely just a visual censor – the site did generate/send the image regardless.
unrelated
And if you have a deviantART account, they have a generator too, though unlike Perchance it always applies a visual style when submitting your prompt. They’ve also recently changed their terms slightly; previously they offered 20 free prompts/month, now it is 10 free prompts/week.
Egad, is that a piece of work (in a good way):
The full prompt
- Prompt
(nude middle-aged female ghost taur:1.5), (see-through body, sleek body, ethereal body, incorporeal, misty body, immaterial, glowing cyan body, cyan skin:1.5), (human face:2), long hair, (two sleek, smooth horns protruding from her head:1.3), expressionless look, hooves, high fantasy, night time, nature background, artstation, pixiv, extremely detailed, standing guard, gripping a longspear with her right hand on its hilt, ready for battle - Negative prompt
solid body, dark skin, (((clothes))), (((nipples))), (((anthro))), ((((ugly)))), (((duplicate))), ((morbid)), ((mutilated)), (out of frame), extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed))), ((ugly)), blurry, ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))), out of frame, ugly, extra limbs, (bad anatomy), gross proportions, (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), (((long neck))) - RNG seed: 480755806
Phrases in parentheses are weighted x1.1 per parenthesis used (alternatively, you can specify a weight directly, as (keyword:weight) )
Style preset used: “Furry - Drawn” (source code)
- Prompt template:
“anthro [_____] illustration, hand-drawn, bold linework, anthro illustration, cel shaded, 4k, fine details, masterpiece” - Negative template:
“bad photo, terrible 3D render, bad anatomy, worst quality, greyscale, black and white, disfigured, deformed, glitch, cross-eyed, lazy eye, ugly, deformed, distorted, glitched, lifeless, low quality, bad proportions”
Note that you prioritized “anthro” as part of your negative prompt, when it was also incorporated into your main prompt by the style preset. This is … not necessarily a problem; repetition does seem to help the AI adjust the weights for certain keywords.
As for me, I made a few runs at Baihu, Brian the Clucky, and Dandy Lion, but none of them yielded anything useful.
1 Like
Blockquote
Note that you prioritized “anthro” as part of your negative prompt, when it was also incorporated into your main prompt by the style preset. This is … not necessarily a problem; repetition does seem to help the AI adjust the weights for certain keywords.
I needed to include it in the negative prompt because otherwise, I often would get an anthro instead of a human upper body because of the art style. And I used the “Furry - Drawn” style because I found it gave the best results for ghostly centaurs.
1 Like
All the style options do is add a list of prompts in both positive and negative lists. You can select the “furry - drawn” option, generate an image and use that image to see what prompts are being used, then input those in and unselect the style. Then you can just remove a single word or phrase rather than trying to add it to the opposing list. You might have to post the image to the gallery and select it from there to see the full lists of prompts, though.
Yeah, I get that. Which is part of why I’m trying to mod a generator to disclose the actual details of the style preset chosen … but this is complicated by all the various import templates that are incorporated, and I don’t know enough about the site’s markup system to get it working correctly … yet.
Presets are awesome when you’re using them as intended, the work they save is real. But sometimes, take even one step outside what it’s designed to do and you end up with MORE work trying to “de-program” what it wants to do…
Anyway, enough rambling. Have a prototype:
Answer and prompt
Working on Clockwork Sphinx.
Prompt: a mechanical sphinx, (quadruped), human face, metal feline body, metal tail, bladed metal wings, clockpunk, digital drawing
RNG seed: 395634317
I tend to have limited patience prompting the AI for anything in specific, because I’m an artist myself: I could spend 3-4 hours making hundreds of tweaks to a text string for the AI to work on, or spend 3-4 hours drawing it myself with the confidence that at least I’ll get it right.
3 Likes
I also got talked into trying Abynissia, but she’s a nightmare. Too many irregular things, some of which I don’t even know how to describe properly in the first place. Anyway, here are some images of her:
4 Likes
Persistence (pun intended):
Prompt used
Prompt
digital painting, upper body and torso closeup, grey humanoid archer, stone wolf statue, angular stone texture, wearing a shoulder quiver and (wielding an archery bow:1.2), ears folded back, snarling, large canine fangs, (armored shoulder plates:1.3) and (arm bracers), glowing blue eyes, narrow gaze, facing right, carved stone, chiseled chest, statuesque
Negatives
armored, shirt, furred, mane, organic
Aside:
- I generally name each result according to its RNG seed (in this case, 669421639). Saves me some hassle of writing it down separately.
- “bow” by itself seems to get interpreted as a decorative bow by default, but what else can I call it?
3 Likes
Harpy Easter!
No GW troops were harmed in creating these images (though the face of the first one reminds me of a … less insane Lyrasza:)
6 Likes
Can we know the name of the artists of the current troops in the game ? I found some but sure there will be others ?
1 Like
Lady Garnetia is that you?
She was born this way:
Prompt: upper body (right side profile) view of a (vivid red:1.2) female anthro (dragon) warrior-knight (walking), flaunting it (like Lady Gaga:0.9), two small red wings, wearing jeweled gold finger claws, gold bracelets and gold shoulder plates, wearing only a (white cloth) brassiere and a black (belt) with gold trim, hourglass figure with (bare scaled midriff), scaled legs and tail, and a gold crown or circlet on her head
Negatives: symmetric composition, robes, dress, skirt, fins
(honestly, it’s not clear what kind of lower body wear Garnetia has, so at some point I gave up trying)
4 Likes
Thank you for reminding me of this thread and for giving me an opportunity to waste some of my time 
3 Likes
Made a few passes at the upcoming Lion Commander, and after maybe 100 iterations or so it produced this:
I couldn’t persuade it to strike the specific pose, so whatever.
Full prompt
(Italics indicate additions from the style preset)
Prompt: anthro yellow lion warrior (with a tail): He wears light armor consisting of a silvery metal breastplate decorated with inlaid (gold trim); shoulder pauldrons expose his furred upper arms while matching silvery bracers are strapped over his forearms. His hands are (clawed) and bestial. A ribbonlike (blue cape:0.95) is attached to this armor, billowing in the wind behind him. His legs are lightly dressed in pants with (leather faulds or tassets) draping down from (the belt). His golden-yellow mane is tied back by a (beaded circlet), with the rest of his bearded hair (tied in braids). He has a confident (smile). His right arm holds a one-edged sword over his shoulder, standing ready to defend the desert city he calls home, digital art, masterpiece, 4k, fine details
Negatives: gloves, gauntlets, bad photo, worst quality, bad composition, bad lighting, bad colors, small eyes, low quality, bad art, poorly drawn, deformed, bad 3D render, boring, lifeless, deformed, ugly, low resolution
1 Like
I like these. Looks much more astral and spirit-y than the original