Continuing previous thread, featuring some upgrades. Finest German engineering.
The AI doesn't understand some names as well as others, so I've taken some liberties to spice up some prompts. But still staying true to the theme of the name.
The base technology is called Stable Diffusion. The model I used for all of these is called Deliberate v2. I'm running it using a tool called AUTOMATIC1111 (technically, "Stable Diffusion Web UI" made by AUTOMATIC1111, but everyone calls it by the author's name). There are other tools which can do similar things, but I like this one since it's one of the more flexible ones. The tool and model are actually capable of much more... what I'm doing to generate these is very basic.
There is a great guide to get started here (also has many useful links beyond that): https://rentry.org/voldy
I'm actually using it via docker, which is a bit different from the above, but I wouldn't worry about that unless you already know what that means. For completeness, here's a link to the dockerized version (which also contains and has links to some alternatives to AUTOMATIC1111): https://github.com/AbdBarho/stable-diffusion-webui-docker
Follow the Voldy guide and you should be up and running in no time. From there, it's one huge rabbit hole after another to figure out all the nuances. But you can do a lot with just the basics.
Not intentionally, I don't think so. It knows Hitler, and Nazi uniforms, etc. Not to mention, there are far more problematic and outright illegal things which it will generate without blinking.
More likely it's a side-effect of AI struggling with details like that. Just like it struggles with hands and writing. It will try to make Swastikas, but they come out garbled. Like it's just not very good at it.
It's a bit counter-intuitive, but a symbol like that requires far more precision than most things. You might think "but it makes such intricate details, like individual strands of hair"... but the thing is, with a strand of hair it doesn't matter if it's exactly there, or two pixels to the left, or three pixels to the right. It'll still look fine as long as the general pattern is consistent. Even with faces, features can be a bit out of place and you'll still recognize the person. But a symbol like a Swastika, if you mess it up even a little, it immediately looks completely broken. I got similarly messed up results when I asked for a Star of David.
Of course, there probably weren't very many labeled Swastikas in the training data, so that doesn't help either.
It's something that can be fixed though. There are LoRAs (kind of like "plugins" for AI models) which let you generate SS Nazi Anime Waifus, for example. Or soI'm told.
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2 yearsApr 29, 2023 15:44:41 ago (+1/-0)
@TheViciousMrPim
Without and with spaces. The former apparently tastes like purple and sunshine.
https://files.catbox.moe/l9fpy2.png
https://files.catbox.moe/f88fyq.png
https://pomf2.lain.la/f/l70kjh3q.png
https://pomf2.lain.la/f/ufxi1hk.png
@BushChuck
This was unexpected...
https://files.catbox.moe/y5tcyt.png
https://pomf2.lain.la/f/ozl9hefy.png
... so I tried adding some terms to make it more relevant:
https://files.catbox.moe/mvrc0g.png
https://pomf2.lain.la/f/7201qf5.png
@big_fat_dangus
NSFL. I'm serious. Don't click it. You've been warned.
https://files.catbox.moe/l9hrrf.png
https://files.catbox.moe/0d5crm.png
https://pomf2.lain.la/f/5j82rf0t.png
https://pomf2.lain.la/f/fe5jabx.png
[ + ] dulcima
[ - ] dulcima 0 points 2 yearsApr 29, 2023 19:42:32 ago (+0/-0)
[ + ] MuricaPersonified
[ - ] MuricaPersonified 0 points 2 yearsApr 30, 2023 15:43:07 ago (+0/-0)
[ + ] lord_nougat
[ - ] lord_nougat 1 point 2 yearsApr 29, 2023 14:24:29 ago (+1/-0)
[ + ] bobdole9
[ - ] bobdole9 1 point 2 yearsApr 29, 2023 12:49:40 ago (+1/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 4 points 2 yearsApr 29, 2023 13:41:47 ago (+4/-0)*
The base technology is called Stable Diffusion. The model I used for all of these is called Deliberate v2. I'm running it using a tool called AUTOMATIC1111 (technically, "Stable Diffusion Web UI" made by AUTOMATIC1111, but everyone calls it by the author's name). There are other tools which can do similar things, but I like this one since it's one of the more flexible ones. The tool and model are actually capable of much more... what I'm doing to generate these is very basic.
You can find the Deliberate model (and many others) here:
https://civitai.com/models/4823/deliberate
AUTOMATIC1111 lives here:
https://github.com/AUTOMATIC1111/stable-diffusion-webui
And has an in-depth wiki here:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki
There is a great guide to get started here (also has many useful links beyond that):
https://rentry.org/voldy
I'm actually using it via docker, which is a bit different from the above, but I wouldn't worry about that unless you already know what that means. For completeness, here's a link to the dockerized version (which also contains and has links to some alternatives to AUTOMATIC1111):
https://github.com/AbdBarho/stable-diffusion-webui-docker
[ + ] Thyhorrorcosmic103
[ - ] Thyhorrorcosmic103 1 point 2 yearsApr 30, 2023 07:06:11 ago (+1/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2 yearsApr 30, 2023 07:16:07 ago (+1/-0)
[ + ] dulcima
[ - ] dulcima 0 points 2 yearsApr 29, 2023 19:47:51 ago (+0/-0)
[ + ] drhitler
[ - ] drhitler 0 points 2 yearsApr 29, 2023 19:05:42 ago (+0/-0)
Seems like the AI really doesn't want to draw a proper Adolf mustache
[ + ] LiberalsAreMental
[ - ] LiberalsAreMental 0 points 2 yearsApr 29, 2023 18:29:53 ago (+0/-0)
[ + ] NaturalSelectionistWorker
[ - ] NaturalSelectionistWorker 0 points 2 yearsApr 29, 2023 18:00:12 ago (+0/-0)
[ + ] SecretHitler
[ - ] SecretHitler 0 points 2 yearsApr 29, 2023 16:23:13 ago (+0/-0)
This is one of the best things I've ever seen I love it
I notice there's no swastikas? Do you think the ai engines are filtered to leave those out?
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2 yearsApr 29, 2023 17:40:33 ago (+1/-0)*
Not intentionally, I don't think so. It knows Hitler, and Nazi uniforms, etc. Not to mention, there are far more problematic and outright illegal things which it will generate without blinking.
More likely it's a side-effect of AI struggling with details like that. Just like it struggles with hands and writing. It will try to make Swastikas, but they come out garbled. Like it's just not very good at it.
It's a bit counter-intuitive, but a symbol like that requires far more precision than most things. You might think "but it makes such intricate details, like individual strands of hair"... but the thing is, with a strand of hair it doesn't matter if it's exactly there, or two pixels to the left, or three pixels to the right. It'll still look fine as long as the general pattern is consistent. Even with faces, features can be a bit out of place and you'll still recognize the person. But a symbol like a Swastika, if you mess it up even a little, it immediately looks completely broken. I got similarly messed up results when I asked for a Star of David.
Of course, there probably weren't very many labeled Swastikas in the training data, so that doesn't help either.
It's something that can be fixed though. There are LoRAs (kind of like "plugins" for AI models) which let you generate SS Nazi Anime Waifus, for example. Or so
[ + ] doginventer
[ - ] doginventer 0 points 2 yearsApr 29, 2023 15:55:38 ago (+0/-0)
You did warn me!
Yeah, the first ones are quite strange enough for me :)
Cheers GB