|
Upscale doing some heavy lifting A milennial of the year 3000  --> 
|
#
?
Aug 17, 2022 00:10
|
|
|
|
| # ? May 30, 2024 20:00 |
|
|
unzin posted:Upscale doing some heavy lifting That looks fantastic.     
|
|
#
?
Aug 17, 2022 01:45
|
|
the Mods, they knew!
|
Well this has turned into quite the loving project. After seeing some others results and what looked like generic Magic cards, I thought it would be kind of fun to reimagine the game's oldest cards (the "Alpha" edition, now almost 30 years old) using prompts describing the artwork of those cards but factoring in the game's now regular artists. I found a few that I pooled together for different things like monsters, art styles, portraits, etc and then got to work starting at the top of the list. Here's the first 45, the white cards in the Alpha deck, now all re-imagined as modern cards:  Lot of work, and this might be the nerdiest thing I've ever done in my life (and that's saying something) but I am having an absolute blast as someone who loves the old cards. It's really interesting to see how sometimes Stable Diffusion does such a great job, and how sometimes it really struggles. I'm not entirely sure it understands the concept of a "pegasus" as something more than a horse with wings attached to it randomly. Going to keep going and will (hopefully) post the results when I'm done along with their original counterparts to compare to. Still, wanted to share my work so far because it's kind of fun to look at.
|
|
#
?
Aug 17, 2022 02:07
|
|
|
ExtraNoise posted:I'm not entirely sure it understands the concept of a "pegasus" as something more than a horse with wings attached to it randomly.
|
|
#
?
Aug 17, 2022 02:09
|
|
|
cross-posting my stuff from the PYF proc gen thread I've gotten Dall-E 2 access. Photograph of master chef making a urinal cake  Photograph of Steamed Hams  A still from the pokemon anime where an ancient evil lovecraftian entity is released  Still of Neil Breen in Dune (1984)  A taxidermied pikachu  Tommy Wiseau in the videogame Silent Hill 4 The Room  Photograph of Cryptid spotted in McDonalds (1996)  A still of Danny Devito in Bloodborne (2015)  A still from the film Bloodborne (2015)  Still of columbo in Bloodborne (2015)  Photograph of eldritch horror working at Wal-Mart  A still of Elmo in A Clockwork Orange (1971)  Still of Guybrush Threepwood in Pirates of the Caribbean (2003)  A still of Kermit the Frog in Vampire Hunter D Bloodlust (2000)  Photograph of a Vampire at a Disco club during the 1970s  Still of Shrek in Avatar (2009) 
|
|
#
?
Aug 17, 2022 02:11
|
|
|
^those are all really dope
|
|
#
?
Aug 17, 2022 02:15
|
|
|
LifeSunDeath posted:oh no doubt, I'm talking about like 3 card monte and poo poo. Or you get a really good scam bot out there setting up MLMs, doing timeshare presentations. its called a slot machine Rutibex fucked around with this message at 02:34 on Aug 17, 2022 |
|
#
?
Aug 17, 2022 02:31
|
|
|
Rutibex posted:its called a slot machiene sure, imagine a slot machine with AI that like became EXTRA addictive as you played! But on 2 legs and it could go out in the community and get me cash.
|
|
#
?
Aug 17, 2022 02:33
|
|
|
LifeSunDeath posted:sure, imagine a slot machine with AI that like became EXTRA addictive as you played! But on 2 legs and it could go out in the community and get me cash.
|
|
#
?
Aug 17, 2022 05:01
|
|
|
Technocloudy With a Chance of Cybermeatballs
|
|
#
?
Aug 17, 2022 15:24
|
|
|
Can someone with access to the good stuff render "neoclassical oil painting of Godzilla eating a cherub" please?
|
|
#
?
Aug 17, 2022 22:33
|
|
|
This Is the Zodiac posted:Can someone with access to the good stuff render "neoclassical oil painting of Godzilla eating a cherub" please?   It's weird, it seems to have issues just drawing Godzilla by itself too. mobby_6kl fucked around with this message at 23:01 on Aug 17, 2022 |
|
#
?
Aug 17, 2022 22:57
|
|
|
This Is the Zodiac posted:Can someone with access to the good stuff render "neoclassical oil painting of Godzilla eating a cherub" please? here's DALL-E out of curiosity 
|
|
#
?
Aug 17, 2022 23:01
|
|
|
I like the way dalle handles this style of prompt for some reason. The strange ways it warps foods are kool   
|
|
#
?
Aug 17, 2022 23:02
|
|
|
Thank you! I like how SD didn't know what a Godzilla was, and Dall-E didn't know what a cherub was 
|
|
#
?
Aug 17, 2022 23:15
|
|
|
This Is the Zodiac posted:Thank you! "godzilla"  "painting of godzilla"  "neoclassical oil painting of godzilla"  And two more attempts at the original prompt   
|
|
#
?
Aug 17, 2022 23:21
|
|
|
mobby_6kl posted:"neoclassical oil painting of godzilla"
|
|
#
?
Aug 17, 2022 23:28
|
|
|
This Is the Zodiac posted:Lower right on this one is fukkin boss. Yeah I've been trying to upscale the bottom two but with no success, the moment you give it different dimensions, it creates a completely different image. That said if you run the prompts individually, it does give you a 512x512 image at least !dream "neoclassical oil painting of godzilla " -S 2498020815 !dream "neoclassical oil painting of godzilla " -S 206892061  
|
|
#
?
Aug 17, 2022 23:40
|
|
|
Can I get a clarification on a thing please...code:code:Correct?
|
|
#
?
Aug 18, 2022 01:52
|
|
|
Yep
|
|
#
?
Aug 18, 2022 01:57
|
|
|
A midjourney attempt. I didn't spend much time tweaking a prompt.
|
|
#
?
Aug 18, 2022 02:45
|
|
|
just made a few things in MidJourney myself   
|
|
#
?
Aug 18, 2022 03:38
|
|
|
Man riding a raft of hamburgers and hot dogs down the Amazon river a giant chilidog floating down the Amazon river 
|
|
#
?
Aug 18, 2022 07:05
|
|
|
a shark leaping out of a sand dune, plume of sand, photograph from national geographic, detailed, 4k
|
|
#
?
Aug 18, 2022 07:17
|
|
|
More MidJourney assisted comics by Abhay Kholsa, this time trying to hide voice over narration as a phone conversation and blending several generated images together.       Compare and contrast with this AI generate image comic that caused a bit of a stir on twitter https://twitter.com/flopscomics/status/1559257365777489920
|
|
#
?
Aug 18, 2022 09:58
|
|
|
what an incredibly myopic view "I guess this could be ok if it helps someone who loses the ability to draw" but I guess it's bad if it helps someone who has never been able to draw? 
|
|
#
?
Aug 18, 2022 10:09
|
|
|
Tracy Morgan in yoda costume and makeup, still from Empire Strikes Back, detailed, 4k
|
|
#
?
Aug 18, 2022 10:17
|
|
|
just saw this in someone else's generated images on stable diffusion I guess there's now a horniness classifier that will blur out your pictures if it thinks they're too sexy from the blurred images it doesn't even look like there was nudity, just near nudity this poo poo better not be baked into the version they release to run on your own hardware, I saw them mention something about making it more "ethical" before releasing it to the public and I was afraid it meant something like this
|
|
#
?
Aug 18, 2022 10:31
|
|
|
TIP posted:what an incredibly myopic view The people who drew and lost it paid their dues and became important to capital in some way, now any cretin can skip that. This is the the danger: money will like you less
|
|
#
?
Aug 18, 2022 10:38
|
|
|
TIP posted:just saw this in someone else's generated images on stable diffusion now it got me  I wanna see the too horny Tracy Morgan TMNT photo they're really doing their best to ruin stable diffusion before it gets out of beta
|
|
#
?
Aug 18, 2022 10:47
|
|
|
Yep in Stable Diffusion I had an image in a series of 4 of the prompt "Biomechanical Pizza Duck by H.R. Giger" get blurred lol Also you can now only generate a maximum of 4 images at a time.
|
|
#
?
Aug 18, 2022 12:29
|
|
|
who do these idiots think they are. the genie is out of the bottle, their efforts will delay porn AI by mere weeks. whoever makes a porno AI first is gonna be a billionare
|
|
#
?
Aug 18, 2022 13:54
|
|
|
I think the devs are treating it as an actual beta test, don't want to get banned from discord and do want to provide a child safe version. We're testing the filter and should report using right click if something is blurred that shouldn't be so they can adjust the filter. Given how libertarian the devs have seemed early on unless someone big is now pressuring them I have a feeling we're getting a local whatever you want version.
|
|
#
?
Aug 18, 2022 13:57
|
|
|
Well their porn classifier is trash, they should probably test that first. Also it's pretty crazy that sometimes it struggles with drawing a dog, but a "cat-caterpillar hybrid" is apparently easy-peasy 
|
|
#
?
Aug 18, 2022 14:12
|
|
|
The NSFW classifier is literally being trained during the beta, which is right now. Moreover they specifically said that they would have it be very harsh to start then walked back. How else are supposed to test this classifier, if not during the beta test? This was all clearly said before hand with cool notifications and everything. This NSFW thing is to be used on their website which they have repeatedly said will be strictly no NSFW. If you use their hardware, no NSFW. If you use your own hardware, say, the local model, no restrictions. This is why the things get blurred instead of simply not existing at all, it stops the results from getting you your eyes, not being generated. The model hasn't changed. And since it's open source plenty of people will probably run porn.ai places that DO let you generate porn. Plus it's been leaked already and 4chan seems to be making plenty of nudity with it. Like poo poo just pay attention, it's a closed discord server that you have to be in to use the thing. They have been very good explaining what they are doing at any time. KakerMix fucked around with this message at 14:31 on Aug 18, 2022 |
|
#
?
Aug 18, 2022 14:29
|
|
|
I assume the filter won't be in the open source version, or will be easy to rip out. It's very unlikely to be baked into the image generation training step, so it would necessarily be another neural network that's run after the first one has finished its work
|
|
#
?
Aug 18, 2022 14:30
|
|
|
ymgve posted:I assume the filter won't be in the open source version, or will be easy to rip out. It's very unlikely to be baked into the image generation training step, so it would necessarily be another neural network that's run after the first one has finished its work It is another AI looking for hotdog, it's likely a simple flag in the full code. Here's some octopus 
|
|
#
?
Aug 18, 2022 14:37
|
|
|
Has anyone figured out what causes the algorithm to freak out and create multiple copies of the same object when it's not supposed to? Seems like a pretty common failure mode, like with Godzilla yesterday or the dogo or diamond joe here.   I stole someone's prompt "anime girl with eyes of the universe, highly detailed, digital art, pixiv fanbox, artstation, by greg rutkowski, wlop, miyazaki hayao" and replaced girl with beagle and that's what it made. KakerMix posted:The NSFW classifier is literally being trained during the beta, which is right now. Moreover they specifically said that they would have it be very harsh to start then walked back. How else are supposed to test this classifier, if not during the beta test? This was all clearly said before hand with cool notifications and everything. This NSFW thing is to be used on their website which they have repeatedly said will be strictly no NSFW. If you use their hardware, no NSFW.
|
|
#
?
Aug 18, 2022 14:43
|
|
|
mobby_6kl posted:Has anyone figured out what causes the algorithm to freak out and create multiple copies of the same object when it's not supposed to? Making the image any size but the default. It can only look at that size area so if it goes to work on another section doesn't see a cat and you have cat in there it's going to make another cat.
|
|
#
?
Aug 18, 2022 14:45
|
|
|
|
| # ? May 30, 2024 20:00 |
|
|
mobby_6kl posted:Has anyone figured out what causes the algorithm to freak out and create multiple copies of the same object when it's not supposed to? Seems like a pretty common failure mode, like with Godzilla yesterday or the dogo or diamond joe here. I assume you are using a different image size than the default 512x512. As far as I can tell, the algorithm for larger canvases is to do a kind of "sliding window" where it starts at 512x512 in the upper left then gradually inpaints into the extended area. The issue seems to be if the object you want is already in the center of the first 512x512 starting area, the inpainting will soon lose track of the object and be like "hey wait the object you wanted is missing!" and add another one. Then it repeats if the algorithm loses track of that object.
|
|
#
?
Aug 18, 2022 14:47
|
|