|
so loving hawt, would rinde.
|
#
?
Sep 3, 2022 18:05
|
|
|
|
| # ? May 30, 2024 04:28 |
|
|
testing out that midjourney thing and its awesome. prompt for this was: stanislaw lem planet of death  planet of death: 
|
|
#
?
Sep 3, 2022 18:10
|
|
|
Boba Pearl posted:I will write you one my friend. Thanks for the detailed instructions! It's good I didn't just try to search that information by myself, because it is quite the laundry list of different ingredients. But I think I'm almost there -- mostly just had to explain to the batch that it shouldn't just look for python, but specifically for python 3.10.6, because the original python installation is an older one from Anaconda.There are still some missing folders I think, but that's probably just a game of patience. I think the script now even auto-downloads some of the dependencies you listed. I'll report back when I'm a bit further along.
|
|
#
?
Sep 3, 2022 20:11
|
|
|
Has anyone been playing around with the model fine tuned on anime? https://huggingface.co/naclbit/trinart_stable_diffusion/tree/main I've been using it for this and that, and I like it, especially for keywords like Akira Toriyama, or Ghibli.
|
|
#
?
Sep 3, 2022 20:28
|
|
|
Rinkles posted:________________ Hey there, I tried reproducing this prompt. I'm getting a beautiful result, but not quite as detailed as the quoted image. Is that to say that the model isn't quite deterministic or was there some post processing like infilling? My result:  
|
|
#
?
Sep 3, 2022 20:31
|
|
|
Are either of you using the optimized version?
|
|
#
?
Sep 3, 2022 20:32
|
|
|
Rinkles posted:
 Yeah stim, you must be using either a different model (This one was generated with 1.4,) or you're using the optimized version for low vram.
|
|
#
?
Sep 3, 2022 20:35
|
|
|
I was checking to see if SD could make a stoner album cover painted by Arik Roper, and it turns out that it can’t. But it’s still pretty funny
|
|
#
?
Sep 3, 2022 20:52
|
|
|
Boba Pearl posted:Yeah stim, you must be using either a different model (This one was generated with 1.4,) or you're using the optimized version for low vram. That was the problem indeed. I wasn't sure which model file to put in the folder initially, so I went with the larger one. Turns out that was the old version. Looks exactly like OP's now. This is so cool!
|
|
#
?
Sep 3, 2022 21:55
|
|
|
When you can train custom modules to load along with the primary one the skill ceiling is going to shoot up as picking which modules to use and which to leave out is going to be very important for results. I expect there will be some public ones that are popular and free, and some paid ones. It doesn't know what a stargate really is, I'm want to train one to make the kawoosh (all things point to training needing more VRAM than I'll have this hardware cycle)
|
|
#
?
Sep 3, 2022 21:57
|
|
|
I'm just waiting for someone to load every frame of the simpsons into a model. There's like 20 million frames of it so even after you prune out the duplicates there's gotta be enough to teach an AI
|
|
#
?
Sep 3, 2022 22:05
|
|
|
Tunicate posted:I'm just waiting for someone to load every frame of the simpsons into a model. There's like 20 million frames of it so even after you prune out the duplicates there's gotta be enough to teach an AI You could probably also get audio description tracks for all the episodes and match the frames to the descriptions so that you have visually labeled reference for the AI.
|
|
#
?
Sep 3, 2022 22:09
|
|
|
Tunicate posted:I'm just waiting for someone to load every frame of the simpsons into a model. There's like 20 million frames of it so even after you prune out the duplicates there's gotta be enough to teach an AI this is gonna be Exhibit A in the trial of humanity
|
|
#
?
Sep 3, 2022 23:04
|
|
|
Trying to make the guy ponder the orb "Bearded old man with hood pondering a big blue orb, magic the gathering art" 30 steps, DDIM, GFPGAN, batch count 2, CFG scale, 512x512, seed 126 Also turns out that Stable Diffusion is absolutely not opposed to drawing, uh, tasteful act portraits for, uh, scholarly pursuits
|
|
#
?
Sep 3, 2022 23:09
|
|
|
Lord Stimperor posted:Also turns out that Stable Diffusion is absolutely not opposed to drawing, uh, tasteful act portraits for, uh, scholarly pursuits "Old man holding blue balls" didn't turn out how you expected eh
|
|
#
?
Sep 3, 2022 23:11
|
|
|
Rutibex posted:"Old man holding blue balls" didn't turn out how you expected eh don't doxx me
|
|
#
?
Sep 3, 2022 23:19
|
|
|
Lord Stimperor posted:Also turns out that Stable Diffusion is absolutely not opposed to drawing, uh, tasteful act portraits for, uh, scholarly pursuits It also does not shy away from adding the occasional surprise third boob (clothed or otherwise), completely unprompted.
|
|
#
?
Sep 3, 2022 23:29
|
|
|
I'm heading to bed, but before I do, I wanted to share with you an exercise I just did. Inspired by the bearded old man with orb, I explored one theme across different parameters and slight prompt variations. I hope this post isn't too long, but I hope that it gives an impression of how a single cool starting image can be shaped. Model version is v1.4. Stable Diffusion: long-haired pale woman with orb (parameters and prompts included in album) I always made batches of two images since that's all my GPU can do. I will try to keep them separate by denoting one of them as [ALTERNATIVE BATCH]. The first images I got were a pretty weird 90s metal album cover, and a vampire fan fiction OC character do not steal image:  Fantasy art, oil painting, long-haired pale woman in a long, white dress, standing before giant orb, crowd in the background Steps: 30, Sampler: DDIM, CFG scale: 11, Seed: 126  [ALTERNATIVE BATCH] Fantasy art, painting, long-haired pale woman in a long, white dress, standing before giant orb Steps: 30, Sampler: DDIM, CFG scale: 11, Seed: 126 The vampire OC do not stel eventually turned into an amazon romance novel portrait:  [ALTERNATIVE BATCH] Fantasy art, intricate, painting, beautiful long-haired pale woman waearing a long, white dress, in a large golden room, facing a giant glowing Orb Steps: 30, Sampler: Euler a, CFG scale: 11, Seed: 126, GFPGAN By describing the background as cosmic, I gave her cool new age vibes (here shown with 60 steps):  [ALTERNATIVE BATCH] Fantasy art, intricate, painting, beautiful long-haired pale woman waearing a long, white dress, in a large golden room, facing a giant glowing Orb, cosmic background Steps: 60, Sampler: Euler a, CFG scale: 11, Seed: 126, GFPGAN On the main image line, I started out with this theme:  Fantasy art, intricate, painting, beautiful long-haired pale woman waearing a long, white dress, in a large golden room, facing a giant glowing Orb Steps: 30, Sampler: Euler a, CFG scale: 11, Seed: 126, GFPGAN More samples and different backgrounds:  Fantasy art, intricate, painting, beautiful long-haired pale woman waearing a long, white dress, in a large golden room, facing a giant glowing Orb Steps: 40, Sampler: Euler a, CFG scale: 11, Seed: 126, GFPGAN By increasing steps you get Stable Diffusion to draw more detailed legs. Galaxy background gives stores and nebulae in the background:  Fantasy art, intricate, painting, beautiful long-haired pale woman waearing a long, white dress, in a large golden room, facing a giant glowing Orb, galaxy background Steps: 60, Sampler: Euler a, CFG scale: 11, Seed: 126, GFPGAN A "cosmic" background gave me ultimately the vibe I wanted, but unfortunately the dynamic pose is gone:  Fantasy art, intricate, painting, beautiful long-haired pale woman waearing a long, white dress, in a large golden room, facing a giant glowing Orb, cosmic background Steps: 60, Sampler: Euler a, CFG scale: 11, Seed: 126, GFPGAN Next I tried out what the CFG scale parameter does. In this motive, it seems to go from abstract, vague to concrete, detailed. I went back to 40 samples since I liked that image, but varied it a little to get some detail in here and there. All images: Fantasy art, intricate, painting, beautiful long-haired pale woman waearing a long, white dress, in a large golden room, facing a giant glowing Orb, cosmic background.  -- CFG 2, steps 40 -- CFG 2, steps 65 -- CFG 3, steps 40 -- these and the next couple of CFG scale are creepy! Skipping. -- CFG 9, steps 40 -- around here, the creepiness stops and we have our familiar theme. -- CFG 10, steps 40 -- it adds some nice tree-like flourishes. -- CFG 11, steps 40 -- flourishes turn to stardust, pretty neat. But pose is off.The alternative image from the batch is actually pretty cool in low CFG scale:  -- CFG scale 1, steps 40 -- neat! -- CFG scale 2, steps 40 -- no magic new age girl, but what looks like a cool biblical angel thing - CFG scale 3, steps 40  -- CFG scale 9, steps 40 -- our magic girl has arrived!
Lord Stimperor fucked around with this message at 00:14 on Sep 4, 2022 |
|
#
?
Sep 4, 2022 00:11
|
|
|
Found this excellent SD prompt and messed around with it, I think its output is insanely good. Replace subject and add attributes like muscular, athletic, obese as needed. Obese (with everything else unchanged) gave me a fully nude output though. Weird. detailed photo of an old bronze patina statue of beautiful lara croft, full body portrait, photorealism, intricate detail, museum diffuse lighting -C 10 -H768    Throwing in Jackman's Wolverine for good measure:  Seems like SD is really good at anatomy. Well, most of the time. (Do I need to NSFW this?) 
|
|
#
?
Sep 4, 2022 02:26
|
|
|
I enjoy that SD knew Logan needed daisy dukes
|
|
#
?
Sep 4, 2022 02:37
|
|
|
mcbexx posted:Found this excellent SD prompt and messed around with it, I think its output is insanely good. I tried to make a bronze fonz and   kinda
|
|
#
?
Sep 4, 2022 02:52
|
|
|
TheWorldsaStage posted:I enjoy that SD knew Logan needed daisy dukes I tried those prompts to make a Daisy Duke statue but no luck with huggingface 
|
|
#
?
Sep 4, 2022 02:53
|
|
|
I closed the Miniconda console before I remembered to copy the whole prompt but I hope you will nevertheless vote #1 Anime golden retriever for president, election_poster, propaganda_poster, celshading, trending on artstation, (some other words)     
|
|
#
?
Sep 4, 2022 04:21
|
|
|
I bought some Dall-E credits to make fantasy arts for a project. This thing is too addictive! I'm keeping the best ones for myself, rare arts only for my book. But please enjoy "carmen sandiego brewing potions by Raphael"
|
|
#
?
Sep 4, 2022 06:00
|
|
|
For those of you running on your own GPU, would a 3080 run much faster than 1080? And any idea how much faster?
|
|
#
?
Sep 4, 2022 07:44
|
|
|
I think I saw somebody saying 6s for 3080 vs 13s for 1080 to render up paintings that never existed. So maybe twice as fast, but I can't recall where I saw that number.
|
|
#
?
Sep 4, 2022 07:49
|
|
|
|
|
#
?
Sep 4, 2022 07:54
|
|
|
Brutal Garcon posted:Very helpful. If you're still in the thread, a new optimized version of Stable Diffusion came out that now does even lower vram. It would be worth setting up now for sure.
|
|
#
?
Sep 4, 2022 08:19
|
|
|
Can you post a link to the new one?
|
|
#
?
Sep 4, 2022 09:36
|
|
|
https://www.reddit.com/r/StableDiffusion/comments/x5dbnj/memoryefficient_attentionpy_updated_for_download/quote:replace the files in stable-diffusion-main\ldm\modules That files it taken from here: https://github.com/neonsecret/stable-diffusion Boba Pearl fucked around with this message at 10:14 on Sep 4, 2022 |
|
#
?
Sep 4, 2022 09:52
|
|
|
That github link still suggests having at least 4gb. Frankly, I should just let this be the thing that finally persuades me to actually spend money on a new computer.
|
|
#
?
Sep 4, 2022 10:24
|
|
|
Boba Pearl posted:I will write you one my friend. Copy -> Paste in Notepad -> Save as txt Very useful info, thanks a lot for that!   
|
|
#
?
Sep 4, 2022 10:29
|
|
|
Bula Vinaka posted:Copy -> Paste in Notepad -> Save as txt Same. Thanks for the foolproof writeup. Brutal Garcon posted:That github link still suggests having at least 4gb. Using the full model on a 3090 with 24 GB still runs out of memory if I try for 1024x1024 images. But 60-step 512x512 ones render in like 3 seconds. Amazing how much latent art, meta-art? is encoded in a binary volume that could fit on a DVD.
|
|
#
?
Sep 4, 2022 11:27
|
|
|
TheWorldsaStage posted:I enjoy that SD knew Logan needed daisy dukes Then you'll probably love this:  I noticed that with a lot of characters (this one, Deadpool, for instance), their color scheme completely overrides the "bronze patina" parameter. How can I put emphasis on that particular part? I tried adding opening/closing parentheses, but that didn't do anything. All Hulks came out green, and not bronze patina green.
|
|
#
?
Sep 4, 2022 12:11
|
|
|
mcbexx posted:I noticed that with a lot of characters (this one, Deadpool, for instance), their color scheme completely overrides the "bronze patina" parameter. How can I put emphasis on that particular part? I tried adding opening/closing parentheses, but that didn't do anything. All Hulks came out green, and not bronze patina green. I think when you give the model a cultural icon it just laser focuses on that and it's incredibly difficult to pry it away from that. Maybe you can add more verbosity with the CFG parameter. Supposedly, that parameter determines how much the model is driven by your exact instructions as opposed to what it infers to fill in the gaps. Another user in this thread was describing how hard it was to stop SD from drawing Mona Lisas. I tried a similar thing: I wanted it to paint a dog in the Girl With a Pearl Earring pose. But the model would invariably just spit out the painting, albeit with the face a bit more hosed up depending on how many instructions I gave it. So I switched my approach and described a golden retriever with a blue hair band, looking over its shoulder, and the model by itself gave it an earring similar to the one in the famous painting:  So certain cultural icons are just super hard to override once it latches on. I would like to figure out how to get around that as well.
|
|
#
?
Sep 4, 2022 12:29
|
|
|
Also a friend and me figured out, uh, for scientific purposes, that Stable Diffusion really likes to draw boobies big.
|
|
#
?
Sep 4, 2022 12:33
|
|
|
How come 1080 ti isn't on there? I'm assuming that would be a bit faster than the normal 1080, right?
|
|
#
?
Sep 4, 2022 12:46
|
|
|
Lord Stimperor posted:I think when you give the model a cultural icon it just laser focuses on that and it's incredibly difficult to pry it away from that. Maybe you can add more verbosity with the CFG parameter. Supposedly, that parameter determines how much the model is driven by your exact instructions as opposed to what it infers to fill in the gaps. Turns out you can actually control the weights with (()) and !!! For lower and higher weights. It's pretty subtle though so you have to really spam it. A Strange Aeon posted:How come 1080 ti isn't on there? I'm assuming that would be a bit faster than the normal 1080, right?
|
|
#
?
Sep 4, 2022 13:02
|
|
|
I have learned that if you want clean line art, you can add the photoshop photocopy filter, and it'll help guide the bot where to put the lines down. It looks pretty good as well.
|
|
#
?
Sep 4, 2022 13:09
|
|
|
|
| # ? May 30, 2024 04:28 |
|
|
a painting depicting the (end of world) and there is [mona lisa] in it in salvador dali style
|
|
#
?
Sep 4, 2022 14:31
|
|