So... anyone been playing with those AI image creators?

[Breakingchains]

'Cause I started fiddling with them as of last night and they're a blast. You feed an AI a text prompt and sometimes a seed or starting image, and using an image database for reference, it does its best to create what you described.

The one I'm using is VQGAN+clip because of its versatility, but neuralblender.com is also popular (faster results and simpler to use, but set dimensions + watermarks on the free version) and I've also given deepdaze a try here and there.

VQGAN+clip has a guide on how to use it here. It takes time to work, but it eventually spits out both still images and a .mp4 of your results developing over time, and does well with more detailed text prompts. Here's what I got when I asked it to create "an old video game about magical adventures and fighting monsters":





...And this is what I got when I plugged in the first few lines of My Immortal (the infamously bad Harry Potter fanfic:)



It ended up with a pretty good idea of her hair, and threw in what might be some band posters, but was completely thrown off the scent by the "Raven" part and seems to have gotten stuck there.

So yeah, it seems like the more specific your prompts, the more interesting the results can be, and it's fun to see what the machine wants to do with different ideas. I'm wondering if anyone else has been playing with these and what are your thoughts?

[Thorn]

How did I not see this before??? I've been mucking around with one of these for just a couple weeks- Artflow, which is specifically for designing humanoid faces. I think it's a lot of fun, and I have friends who use it to determine what d&d characters look like and such. xD

[Breakingchains]

I forgot I made this thread lol

Yes artflow is super cool! I've found it's a little unpredictable because, y'know, machine learning and you never know what sort of associations the system has made, but if you check out their discord server they have a lot of tips to get more consistent results if that's something you want. And it's a ton of fun for experimenting with or quickly iterating on character designs.

It even understands art styles to some degree so you can be like "OK give me an adorable cute disney princess painted by Zdzidslaw Beksinski" so like, have fun with that. xD;

It does seem to strain on nonhuman characters (it understands things like "orc" and "drow" but struggles consistently getting mild nonhuman features like pointy ears, let alone trying to produce e.g. a lizardfolk or something) so considering how many people are using it for DnD and other fantasy characters I'd love to see it get trained specifically on good fantasy art or costumes. Heck, go nuts, chuck a bunch of fursuits in there. Would quickly upgrade it to like my favorite thing on the internet. xD

Actually, on that note, the first time I picked it up, I wanted to test its nonhuman-generating chops, so I punched in "Spongebob Squarepants" and got a very unsettling man with jaundice.

[Killix]

nokemon.eloie.tech/

This one spits out an AI's attempt at generating fake Pokemon.
Like all AI image creators, sometimes you get eldritch abominations, sometimes you get workable designs. XD

[Blueysicle]

Apr 14, 2022 18:48:45 GMT -5 Killix said:

nokemon.eloie.tech/

This one spits out an AI's attempt at generating fake Pokemon.
Like all AI image creators, sometimes you get eldritch abominations, sometimes you get workable designs. XD

Some of the eldritch ones could make for decent Ultra Beasts, at least. xD