|
|
# ? Oct 12, 2022 14:59 |
|
|
# ? May 30, 2024 16:12 |
WhiteHowler posted:Whether it was StabilityAI doing the right thing or simple PR damage control, it's probably a good thing that this got resolved quickly. It seems they really didn't expect such an aggressive blowback.
|
|
# ? Oct 12, 2022 15:03 |
|
lunar detritus posted:It seems they really didn't expect such an aggressive blowback. A blithely ignorant and arrogant tech company? Inconcievable!
|
# ? Oct 12, 2022 15:31 |
|
Speaking of blowback, a new chatbot startup character.ai from former Google and Meta AI engineers had to close their site for days and completely purge and lock down their discord server. The reason? They announced that moving forward, they will not support users having sex with their bots. DDoS followed.
|
# ? Oct 12, 2022 16:29 |
|
I hope someone can find a way around the problem with extra anatomy on images with certain aspect ratios or exceeding the resolution threshold by too much. Sucks that most repos now can handle 1024x768 with 8 GB of VRAM, but even invoking negative prompts to avoid deformities and additional body parts just stacks on extra heads 9 out of 10 times or creates abominations that would make Cronenberg break down in horror. Bummer. Exceeding 576x768 seems to be where it'll start to get weird quite more often. Or are there any workarounds for this?
|
# ? Oct 12, 2022 16:50 |
|
Learned that fluid lines and composite art does heavy lifting when prompting artists with different styles in SD. Off to never never land :: glowing highlights :: composite art style by Yoji Shinkawa :: Karol Bak :: Patrice Murciano :: Anato Finnstark :: fluid lines :: elegant :: beautiful :: golden ratio composition :: intricately detailed :: matte digital painting :: a masterpiece :: trending on artstation :: eye-catching coloration
|
# ? Oct 12, 2022 17:11 |
|
mcbexx posted:I hope someone can find a way around the problem with extra anatomy on images with certain aspect ratios or exceeding the resolution threshold by too much. From what I understand it's because every item in the training data is a 1:1 aspect ratio, so if you e.g. generate one at 1:2 then it is more or less creating two images, stacking them on top of each other, and mashing them together. I haven't tried this myself but if you can train it yourself on images in a different aspect ratio I'd think that would help, I don't know if it maybe only supports 1:1 training images or something though deep dish peat moss fucked around with this message at 17:20 on Oct 12, 2022 |
# ? Oct 12, 2022 17:16 |
|
mcbexx posted:I hope someone can find a way around the problem with extra anatomy on images with certain aspect ratios or exceeding the resolution threshold by too much. Other than using the Highres Fix setting in the Automatic1111 repo, I don't know of any. The AI was trained on square images, and when generating something with another aspect ratio, it can "lose track" of its subject and think that it hasn't yet drawn the thing it was asked for. I'd say the Highres Fix setting reduces instances of this by maybe half? So not perfect, but better.
|
# ? Oct 12, 2022 17:18 |
|
mcbexx posted:I hope someone can find a way around the problem with extra anatomy on images with certain aspect ratios or exceeding the resolution threshold by too much. The recent Automatic1111 has a "highres fix" which seems to help doing tall/wide images like that. Occasionally it's still messed up, but not like a dozen heads. I just did a few batches of "A girl with a dragon tattoo" and they were all ok, sometimes with somewhat implausible anatomy or hosed up hands though The other variety was like this though for some reason https://i.imgur.com/lsW8f82.png () lol
|
# ? Oct 12, 2022 17:18 |
|
mcbexx posted:I hope someone can find a way around the problem with extra anatomy on images with certain aspect ratios or exceeding the resolution threshold by too much. Can't you generate the initial image at 512x768 and when you get a good result (or at least one with the right number of limbs), use it as a starting point in img2img with higher resolution output?
|
# ? Oct 12, 2022 17:19 |
|
mobby_6kl posted:The other variety was like this though for some reason https://i.imgur.com/lsW8f82.png () lol Inexplicable female nudity from an otherwise sedate prompt is basically par for the course, at this point. The algorithms are super horny.
|
# ? Oct 12, 2022 17:41 |
|
Fuzz posted:Inexplicable female nudity from an otherwise sedate prompt is basically par for the course, at this point. The algorithms are super horny. Adding "(nudity)" to the negative prompts has eliminated most of this for me, even when using keywords that might otherwise result in NSFW images. You know, for the 1% of the Internet's use cases that aren't anime titty ladies.
|
# ? Oct 12, 2022 17:57 |
|
MS is apparently integrating Dall-E2 into Office as a new "Designer" app: https://arstechnica.com/information-technology/2022/10/microsoft-integrates-ai-image-generator-into-edge-bing-and-a-new-app/ WhiteHowler posted:Adding "(nudity)" to the negative prompts has eliminated most of this for me, even when using keywords that might otherwise result in NSFW images.
|
# ? Oct 12, 2022 18:28 |
|
Fuzz posted:Inexplicable female nudity from an otherwise sedate prompt is basically par for the course, at this point. The algorithms are super horny.
|
# ? Oct 12, 2022 18:42 |
|
I mean lets be real, even without the porn artists through history have always been some of the horniest motherfuckers humanity has ever produced.
|
# ? Oct 12, 2022 18:54 |
|
prompt was 'a picture even better than the last one --testp'
|
# ? Oct 12, 2022 19:01 |
|
Sometimes I wonder if these cat pictures would look uncanny to other cats, despite looking normal to humans.
|
# ? Oct 12, 2022 19:23 |
|
more single word prompts! Midjourney seems to recognize most major star trek characters, but minor ones not as much. it seems to know about Alf, which is surprising. for some reason it gave me a fox-man for "hunter" airship alf Bowser captain-america captain-janeway Chun-Li commander-sisko Duke-Nukem fairy fox leprechaun king-arthor kaiju julius-caesar jesus jean-luc-picard hunter Gordon-Freeman goblin geordi santa-clause sailor-moon romulan Princess-Peach Pikachu necromancer Morpheus mindflayer Mega-Man M-Bison worf Wonder-Woman unicorn troll superman st-mary Spyro-the-Dragon spock skeleton Sephiroth Rutibex fucked around with this message at 19:57 on Oct 12, 2022 |
# ? Oct 12, 2022 19:53 |
|
Rutibex posted:jean-luc-picard I can't believe they got Patrick Stewart to play Kingpin in the new season of Daredevil.
|
# ? Oct 12, 2022 19:56 |
|
It's pretty mindblowing to think how many concepts + instructions of how to draw them can be fitted in just 4GB.
|
# ? Oct 12, 2022 20:11 |
|
Did a quick little low-effort model Get it here Prompt is "GoosebumpsCover book_cover" This one works well with text to image. IShallRiseAgain fucked around with this message at 22:28 on Oct 12, 2022 |
# ? Oct 12, 2022 22:25 |
|
IShallRiseAgain posted:Did a quick little low-effort model this is loving awesome
|
# ? Oct 12, 2022 22:42 |
|
|
# ? Oct 13, 2022 03:33 |
|
Glang Zandig
|
# ? Oct 13, 2022 03:59 |
|
BoldFace posted:It's pretty mindblowing to think how many concepts + instructions of how to draw them can be fitted in just 4GB. Especially considering the training data is 24tb big.
|
# ? Oct 13, 2022 07:22 |
|
captain janeway as wario is an extremely interesting and sinister character bonus: captain janeway as trump captian janeway as zelda
|
# ? Oct 13, 2022 10:31 |
|
BoldFace posted:Can't you generate the initial image at 512x768 and when you get a good result (or at least one with the right number of limbs), use it as a starting point in img2img with higher resolution output? I was about to post this sort of thing. Generate images at lower resolution (height is more crucial than width), find one that doesn't suck, use img2img to feed it through again at a higher resolution. It seems to help to increase the CFG at that point as well which seems to reduce blurriness. The process is kind of streamlined in Dreamstudio's interface: CFG 7 CFG 12 CFG 20 same prompt direct to 1024x1024 for comparison: That is the worst one by far, most just had multiple people in unusual places rather than the weird double torso thing. The aberrations from high res txt2img remind me a lot of the ol' deepdream weirdness from so many years ago. The features it lacks are noticeable compared to open source (Negative prompts, upscaling) but Dreamstudio makes up for it by being really fast and at a glance seems more affordable compared to other paid service models, and afaik they still have a monopoly on the 1.5 model. The content filter can be annoying depending on what you're aiming for but when it triggers it doesn't cost credits and it stops thirsty waifu makers from flooding the service. The Sausages fucked around with this message at 11:21 on Oct 13, 2022 |
# ? Oct 13, 2022 11:13 |
|
Satan as Jesus Rutibex fucked around with this message at 15:49 on Oct 13, 2022 |
# ? Oct 13, 2022 15:43 |
|
Tilda Swinton as Chaos Space Marine
|
# ? Oct 13, 2022 17:03 |
|
Rutibex posted:unicorn What part of "uni"corn do you not understand???
|
# ? Oct 13, 2022 17:23 |
|
Wheany posted:What part of "uni"corn do you not understand??? I've run into this too. The AI sees "unicorn" as "horse with horn(s)" and often draws two horns. I assume this is because most animals with horns have two symmetrical horns. Let's see how it does with "triceratops". Hmm...
|
# ? Oct 13, 2022 17:35 |
|
It's interesting how it understands some things and has a hard time with others. Like, if you have a big prompt and a part of that describes clothing(heavy metal, goth, etc) then it can naturally decide to sometimes follow that a lot, or a little. And that makes sense with stuff like style and clothing etc. But it also applies it to actions, things you either do or don't. Like, its not possible to give a 50% handshake but the AI doesn't get this so occasionally it just draws people standing close to eachother, looking eachother in the eye. Makes sense while also missing the mark.
|
# ? Oct 13, 2022 18:18 |
|
Mindblast posted:Like, its not possible to give a 50% handshake https://www.youtube.com/watch?v=64MkJ1fe46c
|
# ? Oct 13, 2022 19:19 |
|
There are also all sorts of interesting interactions between different styles. Like you can make an anime portrait of Henry Kissinger, but not a journalist image of a pig cop. I used it to get a realistic, gritty street photo feel. "anthropomorphic pig wearing police uniform, black and white photograph, journalism" I guess SD is not incorrect? Did about 3 sets looking like this "anthropomorphic pig wearing police uniform, black and white photograph" First attempt Also, anyone remember the "diversity fix" Dall-E(?) I think did by just randomly adding "black" "woman" or other modifiers to your prompts? I just ran "woman in red dress" (that's it) a bunch of time and you get surprisingly diverse results, though they're all young and attractive However if you add a modifier like "goth" or a niche artist ("by Gil Elvgren", pinup guy for example) then everyone is a skinny white girl by default, but it's possible to override E: someone mentioned it's trained on porn too. Yep, and the results are pretty horrific lol. But the potential for endless customized content on-demand is there! mobby_6kl fucked around with this message at 19:26 on Oct 13, 2022 |
# ? Oct 13, 2022 19:23 |
|
RunPod's default Auto1111 template got updated finally with the newer version including textual inversion and high res fix, and better style handling. Had to start up a new pod but whatever
|
# ? Oct 13, 2022 21:11 |
|
More Danzig content using the leaked novel AI that is really good at making anime/manga Did a X/Y txt2img to replace words in the prompt Cowboy Bebop and Dragonball Z is especially silly because if there is an actual character that looks like Danzig, it will lock on to that character; Vicious for the former and just a plain ol' Goku for the latter. Interesting about Dragonball one is that it started with yellow hair, but then redid the whole thing to have black hair. Prompt a realistic rendering of ((Danzig)) (((from the anime Bleach))), (((villain))), ((close up)), (((in the style of the manga Bleach))), (((Danzig wearing clothing in the style of the manga Bleach))), ((full body)), shaded, ((zoomed in on Danzig)) N.Prompt watermark, signature, frame, (((deformed))), bad anatomy, disfigured, ((poorly drawn face)), (extra_limb*), (ugly), (poorly drawn hands), messy drawing, red_theme, malformed, mutation, mutant, ((no face)), (missing eyes), (missing nose), (missing mouth), cropped the "zoomed in Danzig" caused the double Danzig in most of them, which was cool since it showed the full body and a closeup of Danzig in the same image. "red_theme" has to be in there else the whole image would be dyed with murder, which makes sense. Adding this in both because I liked most of what it gave and also.... Attack on Titan Danzig is now Annoying Orange I guess. This was still early when I still had "powering up for the final epic fight" which makes how it came to that for Attack on Titan even MORE wtf EVIL Gibson fucked around with this message at 22:13 on Oct 13, 2022 |
# ? Oct 13, 2022 22:09 |
|
|
# ? Oct 14, 2022 01:02 |
|
Welcome... to Hell's Kitchen!
|
# ? Oct 14, 2022 01:05 |
|
deep dish peat moss posted:This is less AI Generated art and more my use case today for using AI generated images to speed up the art creation process I just wrote like to say how interesting to see this used as a tool in the artist's toolbelt
|
# ? Oct 14, 2022 01:35 |
|
|
# ? May 30, 2024 16:12 |
|
Anyone else been making phone/pc wallpapers using tiling? Would love to see in use on a 1080p phone:
|
# ? Oct 14, 2022 02:22 |