Guns are troublesome as well …
Prompt (Captain Macaw)
I like this version best, even if it deviates a bit from the original. He could look a bit more threatening… Well, maybe it’s after the fight 
8 Likes
I threw quite a few random ideas into the generator, but apparently Firefox’s default cookie settings sometimes trigger its scripting for “please don’t adblock” and there were some incidents of the page refreshing automatically, discarding the input prompt (also, results) in the process.
example of a random prompt
On a whim, I input the song lyric “The cold never bothered me anyway” and the generator successfully connected it to Elsa from Disney’s Frozen. Similarly, it mostly correctly read inputs of “Pikachu”, “Charizard” and “Lucario”.
I feel like there’s some kind of blocks around certain stuff, and some things the AI just doesn’t understand. It’s worth trying different words with similar meanings to see if it gets around some kind of block, because there’s no other way to know why the AI won’t do some stuff, or seems to try but doesn’t do it correctly.
I wish it had a feedback system where you could tell it which results were closest to what you wanted, so it would do more like that. You’d think that would help it learn.
1 Like
Geralt of Lawfulgoodia
4 Likes
The Perchance generator definitely seems to have certain things blacklisted … e.g. try to specify that an anthropomorphic subject does not require clothes (or precisely which parts the clothes cover) and it may deem the prompt pretty sus, BUT have no qualms whatsoever about making its own inference that the same subject does not require clothes.
The way these generators “learn” is INCREDIBLY esoteric (the parameter databases range from gigabytes to terabytes and beyond) so I’m not sure it would be of much intuitive help.
Anyway, here’s another go at “Guess That Troop” ™:
The prompt (no presets)
… it was Claw Dancer !
an anthropomorphic cheetah dancer, snarling, wearing tribal earrings and jewelry, arms held wide while wielding daggers (seed:::726945446)
4 Likes
I get very frustrated with trying to create humanoids, but here’s an attempt at Pan’s Vale:
8 Likes
Another take on Claw Dancer. Daggers are ok in this version.
6 Likes
Tried to get it to add the coins along the edges of claw dancer’s outfit, but the AI didn’t seem to want to. Still, got a few more pretty good ones to add.
4 Likes
I am almost there to finish Gemini but i dont understand to prompt and make stained blood on dress. It wont happpen 
If that’s all that’s keeping you from posting it, just post. Someone else might be able to put that blood on there
1 Like
These pics!!! THEY ARE AMAZING!!!
I hope that one day devs will add some of these pics into the game!
@Kafka , @Jeto , @Bramble , maybe you should consider this possibility!
1 Like
We think this thread is really cool for fan art, but we have no plans to add AI art into the game.
Personal opinion: I think AI is cool to play around with for fun and I’ve shared some of these pics with the team as I love this thread and think it’s really cool what you’re all doing. But I’m a firm believer AI should not replace real artists, whose art is used most of the time without permission to train the AI in the first place. Without real artists, what this AI can do wouldn’t be possible and so we should support creative professionals. So for fun, I think everyone should knock themselves out, but as a replacement to a paid artist - absolutely not for me.
Looking forward to seeing what else you all come up with here though
13 Likes
Right now AI can’t replace real artists for most stuff, but in the future it’s going to be competitive, so it’s good for companies to decide now if they want to support artists.
For now, I think it’s a fun tool and it can be used to generate ideas to give to real artists. Like you can input what you want, generate a few pictures, pick the best and give it to the artist, then they can create the ideal version better than an AI can. As long as we use it as a tool, rather than a shortcut, I think it’ll help get better art as a result.
1 Like
My opinion too: I think the prospect of a computer algorithm simulating the work of a human artist AND in only a fraction of the same time is and should be an unsettling one – there is a real danger in what’s known as “technological unemployment”. But also that it’s oddly fascinating regardless to simply watch the process work.
I’ve run quite a few prompts through Perchance’s generator over the past month so here are some observations of my own:
but before that, an aside...
Apparently Perchance.org as a site is all about “generators” – mostly text generators running in a clientside Javascript engine – but they also host a Stable Diffusion setup as a plug-in that people can build/mod their own generators around. Meaning they have literally hundreds of user-created “generators” but all of them ultimately run the same as each other under the hood, the only difference is how the UI organizes the prompt to the end user. For example, one is called a “Fusion Generator” which partitions a prompt into like 7 different fields with various ways of assembling them into the final prompt sent to the generator, but the UI is super complicated so unless you are actually working on a complicated prompt that benefits from being partitioned into multiple groups, it’s really just frontend overkill.
For example, I modded this partly because I couldn’t find a pre-existing generator with the ideal combination of options I was actually using – and I also wanted to keep it “honest” about what it does under the hood.
The AI’s ability to visually render something is limited by its textual interpretation of your prompt. For example, if you give it a prompt of “warrior with a sword” then you could get:
- A warrior with their sword in hand
- A warrior with their sword sheathed
- A warrior standing by a sword stuck in the ground
- A warrior struck by a sword
Similarly, the AI has problems interpreting terms that may or may not embed a specific gender – e.g. “female lion” and “lioness” DO NOT parse the same as each other.
And then there’s the classical problem of ambiguous reads – if given a prompt of “a knight and dragon” then is this read as two characters (one a knight, the other a dragon) or one character (a knight who is also a dragon) ? Likewise, a prompt of “knight fighting dragon with sword” might give the sword to either the knight or the dragon (assuming it got the sword part right to begin with), because the phrasing could technically read either way, and our preference for one read over another is based on cultural knowledge outside the text itself.
A few other observations:
- On average the AI attempts to apply given details across the composition as a whole, meaning that trying to specifically match certain details to certain elements ranges from complicated to impossible. This can create a surprising time-sink where you feel that if the AI can get 90% close to where it should be then why is that last 10% so difficult? Well, because you just don’t have any direct control over the process.
- You thought the AI has trouble drawing hands and fingers? It’s even worse trying to generate text. Because if you want text to be readable then the lettering must be EXACT – there’s simply no room for error here.
2 Likes
Great points. It’s strange when you start to notice how much you don’t know that you know, like being able to know what it likely meant by phrases like “knight fighting dragon with sword” without having to think about it.
It’s also worth noting that certain prompts just aren’t understood the same by us as by the AI. I noticed when I was trying to create a centaur image, I typed in a full description and then kept getting images of humans with horse facial features, rather than actual centaurs, so I eventually simplified it to just “a centaur” and realised the AI doesn’t know what that is, given that it keeps giving horse-faced humans instead of a human body from waist up, horse from the waist down. Knowing that that word won’t work means that I can eliminate it and try others, or give up and try a different troop/idea that might work better.
I’ve found that it helps to have a second page open with the same AI site on it, and if a prompt isn’t working, try putting specific words/phrases in the other page and seeing if the AI gives a good response. If not, you need to change words.
Strange, I didn’t have the same problems:
did somebody say "-taur"s?
prompt: “[___] taur, male” (all using the same RNG seed as each other)
outtakes not shown: fox, otter, dolphin
While some of this absolutely depends on how the AI reads the prompt (e.g. “dragon taur” and “dragon-taur” parsed the same but “dragontaur” did NOT), it seems like more of it depends on the AI’s specific training database (like the ability to recognize certain creature designs by name alone, e.g. “Lucario” or “Protogen”).
1 Like