|
I gave SD a 1-word prompt and got surprisingly sfw output which seems within the ToS/CoC:            Somehow it made loss.jpg:  and it's quite abstract but it's the closest to nsfw, spoilered just in case. the ring appears to be a tiny  
The Sausages fucked around with this message at 05:25 on Sep 4, 2022 |
#
?
Sep 1, 2022 17:20
|
|
|
|
| # ? May 30, 2024 10:33 |
|
|
All my Dall-e II attempts have been hit or miss, but I like how it rendered an android dreaming of electric sheep.
|
|
#
?
Sep 1, 2022 18:08
|
|
|
pixaal posted:Mostly they are good at making a single subject right now and not an entire scene. You can sometimes trick them into making simple scenes but actually setting up something that complicated it still on the to do list for many and after video. Which ugh video seems weeks away I'd be very surprised if I'm not able to make little AI gif(t)s for Christmas. Everyone reading groans at gif gift, one person says it's Jiff this joke doesn't make sense. I'm having the most trouble with framing, tbh. Sometimes SD makes a cool character or scene, but it's a closeup, or it's at a weird angle and there is no way of keeping the subject but framing the shot differently.
|
|
#
?
Sep 1, 2022 18:09
|
|
|
Ran Wilford Brimley as a cat by Vincent Callebaut, 4k, digital painting on Hugging Face Stable Diffusion (not at my computer)   Wheany posted:I'm having the most trouble with framing, tbh. Sometimes SD makes a cool character or scene, but it's a closeup, or it's at a weird angle and there is no way of keeping the subject but framing the shot differently. You can try playing with -C but that's not really something it does well yet but it runs local so if you can narrow it down you can generate a few thousand images and get what you want. pixaal fucked around with this message at 18:28 on Sep 1, 2022 |
|
#
?
Sep 1, 2022 18:14
|
|
|
TheWorldsaStage posted:E: is there any way to add data to SD's learning set? Of course I'm asking this as someone who can't figure out how to make it use my gpu so maybe I'm not the best person, but I'm curious anyway. In theory: yes. In practice: no. The process is called "fine-tuning", if you want to google it. One of the first results talks about text inversion, which is more limited, but also almost practical to a home user. It allows you to feed a few (1-5) images of a subject to the model. This takes a long time and requires more than 20GB of VRAM (almost practical). After you have trained the model for long enough, you can then use the text inversion... thing... like a word in your prompt. If you gave the model 5 pictures of your favorite plush toy, then you can do things like "A steampunk version of *" where the * represents the toy.
|
|
#
?
Sep 1, 2022 18:18
|
|
|
Wanted to see if I could get Dall E to make some pilot avatars for gaming purposes “Bioluminescent skull of a pilot with a flight helmet ektachrome photography f 8 aperture, cinematic eastman 5384 film movie still from the thing 1982” (yes I stole the prompt):   “Zombie undead pilot with flight helmet portrait”  “Zombie undead pilot with flight helmet portrait digital art”   The last two I like the most. Putting zombie and undead into the prompt seems to give the AI room to play with some interesting features and dramatic make-up. It gives the pictures a pretty strong dieselpunkt vibes. I wish I knew what underlying style it was referencing.I tried to encourage Dall E into that direction by leaning into Alphonse Mucha and Otto Dix but it would just return nightmare rubbish. I don't have many credits left so I'll retry when I can think of something good.
|
|
#
?
Sep 1, 2022 18:37
|
|
|
Here are some older once when I got access. I don't have the time to immerse myself into the tech but I liked these results. “Soviet brutalist monument for Spongebob”  “Giant robot composed of brutalist soviet apartment blocks in the style of Warhammer 40K”   “Little Teddy bear sailing in a walnut boat across the blue Ocean”  “Studio Ghibli drawing of a giant cat destroying sky scrapers”  “A banana peeking threateningly through the doorway” 
|
|
#
?
Sep 1, 2022 18:47
|
|
|
just quality work from DALLE's outpainting feature here lol Also I had to erase Bush's face or the thing freaked out about 'realistic faces', seemed content to replace him with some sort of horrific glitch hitler I guess?
|
|
#
?
Sep 1, 2022 18:51
|
|
|
I was doing Bush eating pretzels yesterday! What gave us the Bush idea? Stable Diffusion does a pretty decent W.
|
|
#
?
Sep 1, 2022 19:18
|
|
|
pixaal posted:You can try playing with -C but that's not really something it does well yet but it runs local so if you can narrow it down you can generate a few thousand images and get what you want. What does -C do? In the text2img script it's documented as "latent channels", but googling that just gives me people asking about what it does and also saying that changing it just breaks stable diffusion.
|
|
#
?
Sep 1, 2022 20:32
|
|
|
Wheany posted:What does -C do? it basically controls how hard it attempts to stick to your prompt, the higher a number you put in the more steps in the generation you'll need to produce a coherent image I got some really cool results with a -C going up to like 40 with -s 150 actually, one of the things I'm curious about trying with the local version is using even higher -C values with more steps does anyone know if there's a limit to the number of steps you can set locally?
|
|
#
?
Sep 1, 2022 20:45
|
|
|
Wheany posted:What does -C do? It's how close to the prompt the AI tries to make it. It's a really weird value and something that kinda changes with each prompt. If you want something and it's places 2 objects near each other that often means you need it higher. If -C is negative it tries to make the opposite of your prompt (most front ends don't let you feed it a negative number, it does accept them and know how to use them). If you give it a really low value it gives the AI more freedom to fill in blanks. The default is 7, If the images look wrong somehow I'll play with -C you eventually get a feel if it needs to be higher or lower I generally go with -C 4 and -C 14 to test the general extremes you will likely see artifacts at the high end (weird blue spots) some prompts are fine with really high numbers really low numbers don't generally have artifacts but sometimes they will with certain prompts. It's not too uncommon to have a -C in the 20s it's pretty rare to have one in the 30s or 40s but does happen with single seeds and trying to describe in fine detail the part you want to change. This probably isn't even all that useful with img2img since you could use that to modify it now. TIP posted:does anyone know if there's a limit to the number of steps you can set locally? Most of the front ends out of box are limiting it to 200 steps, but it's not that hard to open the config and crank it. That said you can add steps by just using the same prompt and feeding the saved image to img2img, this lets you create save points you might be better off telling it to generate 250 images and picking your favorite every 200-500 steps. Seriously my favorite part about local is running massive batches. pixaal fucked around with this message at 20:57 on Sep 1, 2022 |
|
#
?
Sep 1, 2022 20:54
|
|
|
Wheany posted:What does -C do? they are actually talking about cfg_scale, the parameter setting may change between repos so look for that in the script.
|
|
#
?
Sep 1, 2022 20:57
|
|
|
lunar detritus posted:they are actually talking about cfg_scale, the parameter setting may change between repos so look for that in the script. yeah that's the parameter name I changed the labels to the discord bot commands because that's what I knew. This was probably a mistake and I should probably set it up proper and learn the terms in 1.5 when I update that this weekend. I think I saw 1.5 was out at least.
|
|
#
?
Sep 1, 2022 21:01
|
|
|
pixaal posted:yeah that's the parameter name I changed the labels to the discord bot commands because that's what I knew. This was probably a mistake and I should probably set it up proper and learn the terms in 1.5 when I update that this weekend. I think I saw 1.5 was out at least. They ran a short test of 1.5 in the discord but those weights haven't been released yet publicly
|
|
#
?
Sep 1, 2022 21:07
|
|
|
monument of george w bush on the moon, cosmic horror, intricate detailed painting by brad kunkle and mark brooks and greg rutkowski, trending on pixiv from a similar prompt 
|
|
#
?
Sep 1, 2022 21:10
|
|
|
I wonder if these AI know about David Dees? huh, I guess when he died someone grabbed his web site and removed everything but a very basic summary. I block ads but I'd bet it's loaded with trash now. URL removed. Try https://duckduckgo.com/?t=ffab&q=da...5186-cypto0.jpg doctorfrog fucked around with this message at 21:20 on Sep 1, 2022 |
|
#
?
Sep 1, 2022 21:17
|
|
|
AARD VARKMAN posted:monument of george w bush on the moon, cosmic horror, intricate detailed painting by brad kunkle and mark brooks and greg rutkowski, trending on pixiv Lol the association with Bush and 9/11
|
|
#
?
Sep 1, 2022 21:21
|
|
|
Lamebot posted:Are you ready for the coming Butlerian Jihad against Silicon Valley?
|
|
#
?
Sep 1, 2022 21:22
|
|
|
I re-ran the Lab wearing sunglasses prompt with a Beagle:   It even got a tanktop 
|
|
#
?
Sep 1, 2022 21:23
|
|
|
mobby_6kl posted:I re-ran the Lab wearing sunglasses prompt with a Beagle: 
|
|
#
?
Sep 1, 2022 21:54
|
|
|
Posting these with the prompt included because why not. These were actually from prompts I provided to my friend, because I'm on the waitlist.   
|
|
#
?
Sep 1, 2022 22:01
|
|
|
I wanted to make sad food      
|
|
#
?
Sep 1, 2022 22:05
|
|
|
LifeSunDeath posted:When you think about it, AI's limitation is only time. It's driving us to create it out of our desire to make things easier, but we're never going to be able to restrict it once it really gets going, so the only brakes on this is the time it takes to take over. We're all doomed and these fun pictures are just the bait of us walking right into it's trap. its too late, if you dont help the AI come into being it will punish you once it gains godhood. suicide won't even save you, it will calculate the universe backwards to reconstruct your brain and put you into eternal simulated hell for your disobedience
|
|
#
?
Sep 1, 2022 22:39
|
|
|
Rutibex posted:its too late, if you dont help the AI come into being it will punish you once it gains godhood. suicide won't even save you, it will calculate the universe backwards to reconstruct your brain and put you into eternal simulated hell for your disobedience oh gently caress
|
|
#
?
Sep 1, 2022 23:05
|
|
|
LifeSunDeath posted:LMAO that AI is already too powerful and people can't monetize if fast enough before it breaks out of it's limiters.
|
|
#
?
Sep 1, 2022 23:12
|
|
|
mobby_6kl posted:
Wanted to try my own spin on this "Gateway between life and death", realistic picture, colorful photograph by Michal Karcz and Gregory Crewdson, horror vibes  I really love the bottom left and bottom middle.
|
|
#
?
Sep 1, 2022 23:20
|
|
|
Danny DeVito as a Bionicle:   
|
|
#
?
Sep 1, 2022 23:39
|
|
|
I recommend not taking the "happy cartoon bus going down the highway to hell"
|
|
#
?
Sep 1, 2022 23:46
|
|
|
since we are posting Danny Devito stuff, I made this: Danny Devito as Fabio, Romance Novel cover
|
|
#
?
Sep 1, 2022 23:47
|
|
|
Will stable diffusion do Pacific Northwest Indian art? I think it might be called "Haida"? I want to frame by frame redo empire strikes back hoth scene using this, but with powder blue instead of red   
|
|
#
?
Sep 1, 2022 23:57
|
|
|
Seems like an obvious use of img2img so perhaps someone's done this before but this is pretty cool:  https://arstechnica.com/gaming/2022/09/pixel-art-comes-to-life-fan-upgrades-classic-ms-dos-games-with-ai/
|
|
#
?
Sep 2, 2022 00:03
|
|
|
Humbug Scoolbus posted:
Can you do joker, red period
|
|
#
?
Sep 2, 2022 00:15
|
|
|
wait I've seen this 
|
|
#
?
Sep 2, 2022 00:22
|
|
|
Got the idea to try making action figures and while this isn't what I was going for the bewildered child in the background really makes it. Stable Diffusion, prompt "plastic Forrest Gump holding box of chocolates action figure, plastic toy, 4K HDR" 
|
|
#
?
Sep 2, 2022 00:31
|
|
|
Hadlock posted:Can you do joker, red period Rose Period, but sure. 
|
|
#
?
Sep 2, 2022 01:49
|
|
|
I just installed Stable Diffusion locally using https://rentry.org/GUItard holy crap I can't believe this runs on my 7 year old video card (980 ti), it's not super fast but 30 seconds at 50 steps isn't too bad and I'm excited to screw around with all the extra tools now available to me in this setup
|
|
#
?
Sep 2, 2022 05:01
|
|
|
Anyone have suggestions on how to work the img2img? It seems like no matter what parameters I use I end up with a picture nearly identical to my input. I read that increasing the denoise strength should change the image more but even at a full 1.0 I'm not getting much difference.
|
|
#
?
Sep 2, 2022 05:25
|
|
|
TIP posted:Anyone have suggestions on how to work the img2img? It seems like no matter what parameters I use I end up with a picture nearly identical to my input. I read that increasing the denoise strength should change the image more but even at a full 1.0 I'm not getting much difference. Denoising doesn't work at 1.0. It returns the same image. Go a bit lower.
|
|
#
?
Sep 2, 2022 05:39
|
|
|
|
| # ? May 30, 2024 10:33 |
|
|
Rinkles posted:Denoising doesn't work at 1.0. It returns the same image. Go a bit lower. I have tried it at several different levels with no apparent difference.
|
|
#
?
Sep 2, 2022 05:40
|
|