|
PlaysGamesWrong posted:
Chairman Maovito
|
#
?
Apr 18, 2023 21:28
|
|
|
|
| # ? May 30, 2024 19:02 |
|
|
deep dish peat moss posted:e: In short, the fears that the entire internet will be a sea of AI-generated drivel are entirely legitimate, and it's only going to get worse when advertisers start spinning up grassroots marketing accounts on Social Media and Reddit et al with discrete purpose-built AI designed to "prove" that they are human and not AI, at some point even all of the "people" that you meet online are likely to be AI hell-bent on proving to you that they're not AI so that they can subtly sell you things. We really should have an AI chat thread at some point, this has been the catch all thread but when the copyright office is getting involved at the federal level it's probably worthy of a distinct thread Anyways I kind of think it's going to be the opposite, ChatGPT free edition is already a drop in replacement for 99% of random Google searches "when should I plant tomatoes in new York?" Or "what are the best seven ski areas within a 10 hour drive of San Francisco?" you don't need a search engine or to read six computer generated blog articles to find out the answer to that. ChatGPT will just give you the straight up data like a personal wikipedia researcher or something I think it's going to be the complete opposite, 95% of trash, autogenerated blogs will just vanish and the weird ad supported niche they filled will evaporate over 5-10 years. ChatGPT is just going to be filled with ads for gardening tools and ski resorts Google already announced they're writing a new "search engine" from scratch that will basically be ChatGPT but with ads. Microsoft already has a 18 month head start in the topic Maybe even more democratizing is the fact that anybody with a million bucks can train their own LLM, there's a truly open source model called red pajamas (take on LLM and llama rhyming scheme) and I'm sure it won't be the last. Googles economic moat was that operating at their scale was impossible. Now anyone with a 3090 GPU or $10,000 in AWS credits can become a credible competitor I'm not gonna come out and say Google's hosed, but they've certainly perked up in the last 12 months TL;DR not only is it better but it's cheaper too
|
|
#
?
Apr 18, 2023 22:06
|
|
|
messed around with the Web-based Starry.ai a little today, I don't care for the interface but here's what I enjoyed photo: multicolored tardigrade eating a volvo  art: Arabic calligraphy detailed VCR manual   art: Arabic calligraphy volleyball club logo 
|
|
#
?
Apr 18, 2023 22:13
|
|
|
KwegiboHB posted:No we will not and I will fight that for the rest of my life. You will never EVER get that back once it's gone. Ever. good luck! I think it's inevitable, I'd love to be wrong.
|
|
#
?
Apr 18, 2023 22:30
|
|
|
Gynovore posted:This isn't a binary choice, it's not a Ready Player One style Evil Corp vs. everyone thing. You can have the "brightly lit' internet alongside the current "wild west" implementation. we could also have both public and private healthcare, and yet
|
|
#
?
Apr 18, 2023 22:31
|
|
|
Nobody cares. MJ and I have a difference of opinion about what constitutes a "biblically accurate angel" but the biblically accurate toaster ones are at least amusing.   
|
|
#
?
Apr 18, 2023 22:33
|
|
|
BARONS CYBER SKULL posted:basic workflow! thanks again for this; the image here is nothing I'm super attached to but the overall process is great and the resolution / detail added at each step here was cool and I look forward to tweaking this more (and also seeing what settings I can get away with on a 4090, what are you using?) Here's original (highres 2x) image, then 1st pass (~1 min), then 2nd pass in full res (~6 mins). For giggles I have fired off another pass which says it is 484 tiles (compared to 121 tiles for pass 2), so, curious to see if that completes.   
|
|
#
?
Apr 18, 2023 23:08
|
|
|
What does a biblically accurate lobster trauma surgeon look like
|
|
#
?
Apr 18, 2023 23:34
|
|
|
Hadlock posted:What does a biblically accurate lobster trauma surgeon look like I don't think MJ really knew what to do with it 
|
|
#
?
Apr 18, 2023 23:45
|
|
|
Hadlock posted:What does a biblically accurate lobster trauma surgeon look like Did you mean a trauma surgeon for lobsters, or a lobster who is a trauma surgeon? Doesn’t matter, it has you covered! 
|
|
#
?
Apr 18, 2023 23:45
|
|
|
LOL beaten. When I went to MJ I saw I had typed in “I stink like fish” and let it run. Can’t for the life of me remember why, but I REALLY like this one. I may have T-Shirts made. 
|
|
#
?
Apr 18, 2023 23:47
|
|
|
Cabbages and Kings posted:At some point we're going to need the super locked-down, corporate internet that most of us have been dreading because the open internet is going to be nothing but AI generated political disinfo and AI generated explicit content. you do realize that most of the current political disinfo is already shared on super locked-down corporate systems like Facebook? corporate control won't "save" us, and we should not desire it.
|
|
#
?
Apr 18, 2023 23:48
|
|
|
Doctor Zero posted:Did you mean a trauma surgeon for lobsters, or a lobster who is a trauma surgeon? Doesn’t matter, it has you covered! I was thinking something like Dr zoidberg  that bottom right one is roughly in line with my mental image I hadTop left here also a winner: PlaysGamesWrong posted:I don't think MJ really knew what to do with it
|
|
#
?
Apr 18, 2023 23:52
|
|
|
Yeah I'm not really understanding that, the Corp internet is what has made the internet suck with it's engagement SEO algorithms. Corps made the internet worse because now there is profit in misinformation. Before it was bigfoot and I believe flying saucers. Now it's pizza gate and qanon. Corporations did that, Fox news is that.
|
|
#
?
Apr 18, 2023 23:53
|
|
|
 https://files.catbox.moe/zf69r9.png blew that up again, it's a 75mb file so I am not hotlinking, but as an example of the detail, here is the face above the head of the meditating woman  this is completely bonkers, can just make loving 8k wallpapers lickity split. Going bigger via webUI seems sketchy just because the file sizes here already cause chrome to choke, but if "generate" is jammed and completes in the shell then the image has probably been written to disk even if it kills the browser. I didn't even know imgur had a limit, or what catbox is, but here we are. Hadlock posted:What does a biblically accurate lobster trauma surgeon look like 
|
|
#
?
Apr 18, 2023 23:55
|
|
|
Cabbages and Kings posted:thanks again for this; the image here is nothing I'm super attached to but the overall process is great and the resolution / detail added at each step here was cool and I look forward to tweaking this more (and also seeing what settings I can get away with on a 4090, what are you using?) just a 2080 Super here also an update on the workflow, on the second pass (so after you drag over the upscaled img2img one), drop the denoise down to 0.2, or lower, and you'll get rid of the multiple faces stuff that happens, i think i forgot to mention that
|
|
#
?
Apr 19, 2023 00:15
|
|
|
BARONS CYBER SKULL posted:just a 2080 Super here Thanks. Weirdly a lot of that stuff took care of itself in the upscale to the final 75mb file, and also I think I had "cyber church background, worshippers" in the prompt which may add ... Entities.
|
|
#
?
Apr 19, 2023 00:23
|
|
|
 
|
|
#
?
Apr 19, 2023 00:25
|
|
|
neat. old sketch I did, again:  substantial upscale:  the upscale after this went to 45mb
|
|
#
?
Apr 19, 2023 01:19
|
|
|
Oh neat, I forgot who yoshitaka-amano was, shout out to whoever clued me into the url hints. I had no idea I could make final fantasy styled lobster trauma surgeons  No idea why but mid journey seems to make the best photos for whatever reason
|
|
#
?
Apr 19, 2023 01:36
|
|
|
Cabbages and Kings posted:good luck! I think it's inevitable, I'd love to be wrong. Let's talk of many things, why is the sea boiling hot? Do pigs have wings? The good news is, you are not alone in the universe. In all this talk of AI it's possible to lose sight of the 8,000,000 I that already exist. Some of us want all of us to bloom. The Smithsonian just put 4.5 million Hi-Res images into the Public Domain. Here's a huge list of tools and resources to use them with It doesn't all have to come from one person or source though, adding up a thousand small helpful gestures, bits of advice, sharing and pooling of knowledge, results of failures as examples to learn from, all of these make the world a bit brighter than it was the day before. I'll be working on my own Stable Diffusion guide later even though I know the tools I use today are going to look nothing like the tools I use a year from now. I have no timetable but eventually I'll also make a full custom Stable Diffusion model trained entirely off my own work. I know that will be quite popular so I'll make sure to write guides about that as well. I'm really looking foward to having a full Home AInimation Studio. boooooo It doesn't take full on galaxy brain to have a positive impact, mostly it's just writing down what you do and sharing it for others to learn from. Like this excellent guide but with a twist. BARONS CYBER SKULL posted:basic workflow! I threw a base image in...  Maybe poo poo doesn't have to be so hosed up yo. We could end up with a Golden Age out of this instead of continued dystopia. SolarPunkAI <lora:solarpunkai_v10:1> Negative prompt: Give all your money and rights to billionaires. Steps: 64, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 3596134797, Size: 512x512, Model hash: 7f16bbcd80, Model: dreamshaper_4BakedVae, Denoising strength: 0.69, Clip skip: 2, Hires upscale: 1.5, Hires upscaler: Latent hi-res fix version...  Maybe poo poo doesn't have to be so hosed up yo. We could end up with a Golden Age out of this instead of continued dystopia. SolarPunkAI <lora:solarpunkai_v10:1> Negative prompt: Give all your money and rights to billionaires. Steps: 64, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 3596134797, Size: 512x512, Model hash: 7f16bbcd80, Model: dreamshaper_4BakedVae, Denoising strength: 0.69, Clip skip: 2, Hires upscale: 1.5, Hires upscaler: Latent upscaled for extra details...  Maybe poo poo doesn't have to be so hosed up yo. We could end up with a Golden Age out of this instead of continued dystopia. SolarPunkAI <lora:solarpunkai_v10:1> Negative prompt: Give all your money and rights to billionaires. Steps: 64, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 1484508584, Size: 1024x1024, Model hash: 7f16bbcd80, Model: dreamshaper_4BakedVae, Denoising strength: 0.69, Clip skip: 2, Mask blur: 4, Tiled Diffusion upscaler: 4x-UltraSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 96, 'Latent tile height': 96, 'Overlap': 48, 'Tile batch size': 1, 'Upscaler': '4x-UltraSharp', 'Scale factor': 2, 'Keep input size': True}" then re-shrunk back down to 512x512 and thrown in another round for Extra extra details!  Maybe poo poo doesn't have to be so hosed up yo. We could end up with a Golden Age out of this instead of continued dystopia. SolarPunkAI <lora:solarpunkai_v10:1> Negative prompt: Give all your money and rights to billionaires. Steps: 64, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 2018020093, Size: 2048x2048, Model hash: 7f16bbcd80, Model: dreamshaper_4BakedVae, Denoising strength: 0.69, Clip skip: 2, Mask blur: 4, Tiled Diffusion upscaler: 4x-UltraSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 96, 'Latent tile height': 96, 'Overlap': 48, 'Tile batch size': 1, 'Upscaler': '4x-UltraSharp', 'Scale factor': 2, 'Keep input size': True}" I don't see a ceiling to this yet, weeeeeeeeee. If anyone with a 4090 wants to try a few recursive detailed iterations I'd appreciate it because this is sloooow on my machine. Oh and while I'm on my soapbox, I really need to stress this for those that need to hear it, Stable Diffusion is two separate parts. The software that runs the Diffusion Process is distinct from the Model Card trained from whatever dataset. I can understand someones eyes glazing over when they view 'explanations' like these but one basic part is consistently left out: You can just plop a new model in anytime and get completely different results. In all this rush to figure out what's going on I need to make sure the VCR player isn't banned because someone didn't like what was on a VHS. The cassette player because someone didn't like a tape. The record player because someone didn't like an album. Call your grandparents and tell them about that part, they miss you anyways. Edit: timg'd the 2k picture. KwegiboHB fucked around with this message at 02:26 on Apr 19, 2023 |
|
#
?
Apr 19, 2023 02:03
|
|
|
gently caress y'all I feel old. All this talk of images and memes made me want to go check in on some of the classic memes I remember from early in my career, and god drat the sands of time hath not always been kind.     ymgve posted:you do realize that most of the current political disinfo is already shared on super locked-down corporate systems like Facebook? lol yes I very much realize that, I've spent most of my career as a cog in the growing panopticon, it was intended as doomer snark, nothing more. KwegiboHB posted:I don't see a ceiling to this yet, weeeeeeeeee. If anyone with a 4090 wants to try a few recursive detailed iterations I'd appreciate it because this is sloooow on my machine. it's "sealing", like "sealing wax". But I will do whatever you would like with a 4090 if you give me a source image and very precise A-B-C steps, because my brain is fried from mode switching programming languages lately and in my free time I've been acting like a thorazine patient. Cabbages and VHS fucked around with this message at 02:11 on Apr 19, 2023 |
|
#
?
Apr 19, 2023 02:08
|
|
|
BARONS CYBER SKULL posted:also an update on the workflow, on the second pass (so after you drag over the upscaled img2img one), drop the denoise down to 0.2, or lower, and you'll get rid of the multiple faces stuff that happens, i think i forgot to mention that Oh yeah, that's quite a bit better.  Maybe poo poo doesn't have to be so hosed up yo. We could end up with a Golden Age out of this instead of continued dystopia. SolarPunkAI <lora:solarpunkai_v10:1> Negative prompt: Give all your money and rights to billionaires. Steps: 64, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 1380485857, Size: 2048x2048, Model hash: 7f16bbcd80, Model: dreamshaper_4BakedVae, Denoising strength: 0.2, Clip skip: 2, Mask blur: 4, Tiled Diffusion upscaler: 4x-UltraSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 96, 'Latent tile height': 96, 'Overlap': 48, 'Tile batch size': 1, 'Upscaler': '4x-UltraSharp', 'Scale factor': 2, 'Keep input size': True}" Cabbages and Kings posted:gently caress y'all I feel old. lol, these are amazing. I was just following BARONS CYBER SKULLs guide except i shrunk the output back to 512x512 and re-ran it one more time hoping for more details to appear out of nothingness. I see now that will depend on the de-noising steps but it takes me 20 minutes to make one of these so A/B testing will take me a bit to find a sweet spot. It's less the original image that's put in, than where and what extra details appear from nothing. Feel free to try anything you like.
|
|
#
?
Apr 19, 2023 02:34
|
|
|
 https://www.bbc.com/news/entertainment-arts-65296763
|
|
#
?
Apr 19, 2023 03:20
|
|
|
Yeah, we know.
|
|
#
?
Apr 19, 2023 03:57
|
|
|
Just dicking around and realized that you can write prompts in other languages and even mix them up. Like if you use the Japanese word for woman you get a woman more likely to appear Asian. I feel like there might be some good magic words in other languages nobody has found yet, like tied to what people in China or something might tag their photos that doesn't translate similarly to English. Idk. Any polyglots in the house?
|
|
#
?
Apr 19, 2023 05:01
|
|
|
Sometimes I gently caress up webui's memory state to where I need to kill it and restart, which is annoying if I'm using it from the couch over the network. Switching models seems to be the worst offender. There's no hidden lever to fully reload, is there? I know the UI can be reloaded from the extension menus, but I don't think that has any impact once it's in a broken state (but still responsive to loading the UI)
|
|
#
?
Apr 19, 2023 05:30
|
|
|
PlaysGamesWrong posted:Just dicking around and realized that you can write prompts in other languages and even mix them up. Like if you use the Japanese word for woman you get a woman more likely to appear Asian. lmao what are you even looking for?
|
|
#
?
Apr 19, 2023 05:34
|
|
|
Cabbages and Kings posted:There's no hidden lever to fully reload, is there? I don't think so? You can right click webui-user.bat and send a shortcut to desktop, rightclick the shortcut and set a shortcut key like f10. Then you can just start or restart the whole thing with a button press. You'd still have to close the frozen model before.
|
|
#
?
Apr 19, 2023 05:54
|
|
|
PlaysGamesWrong posted:Yeah, we know.
|
|
#
?
Apr 19, 2023 07:02
|
|
|
It's a bananaoff! Let's see what the free ai tools make of this prompt: "A regular normal banana on a plain white background. Soft lighting. Color photo." Craiyon  4/9 are passable as a regular banana. I like the loopback banana and the one that has the stem sticking out at a right angle. Stable Diffusion 1  Maybe 1/4? The last one looks more like a jalapeño with banana colors. Stable Diffusion 2.1  No. Not a single normal regular banana in the bunch. And the winner? Dall-E  Look at those boring bananas. Good job, I guess?
|
|
#
?
Apr 19, 2023 10:00
|
|
|
PlaysGamesWrong posted:Just dicking around and realized that you can write prompts in other languages and even mix them up. Like if you use the Japanese word for woman you get a woman more likely to appear Asian. ChatGPT
|
|
#
?
Apr 19, 2023 10:13
|
|
|
KwegiboHB posted:I don't think so? You can right click webui-user.bat and send a shortcut to desktop, rightclick the shortcut and set a shortcut key like f10. Then you can just start or restart the whole thing with a button press. You'd still have to close the frozen model before. Sure, I just use it over the network sometimes and that PC is cut presently set up for easy full remote access.
|
|
#
?
Apr 19, 2023 12:14
|
|
|
loving with this stuff on a regular basis has created a bizarre new kind of writer's block for me, like, "what do I want to see today?" This morning's effort has been focused on quote:a beautiful smiling cat in a purple top hat standing outside a mcmansion in a suburban development, middle america background, sunny day, vivid, extremely detailed Why? What part of my brain wanted to put those pieces of reality together in that way? 
|
|
#
?
Apr 19, 2023 13:30
|
|
|
PlaysGamesWrong posted:Just dicking around and realized that you can write prompts in other languages and even mix them up. Like if you use the Japanese word for woman you get a woman more likely to appear Asian. I haven't tried this in MJ, but I will. I have done this in ChatGPT and its language ability is pretty rudimentary. <Language> to English is pretty good. English to <Language> is not any better than Google Translate, though.
|
|
#
?
Apr 19, 2023 14:38
|
|
|
ChatGPT's translation is useful when you want to actually do something else along with translating the text. We used it to write Russian firmware for a machine that originally had Chinese firmware, and so had a character limit that was more in line with Chinese words. We needed to translate the Chinese and abbreviate it in a way that still made sense to a Russian person (and wouldn't have if we were just chopping up Russian words without understanding any Russian.) Like, it would find smaller synonyms when possible and do common abbreviations for the words. We double checked it with a real Russian person and they were like "ya I can read this" Originally it wasn't gonna be used, but I just saw a video of it in action and I guess they went ahead with it. Which made me feel really smart cuz it was my idea.
|
|
#
?
Apr 19, 2023 14:51
|
|
|
Megazver posted:ChatGPT  LLMs are generally pretty good with getting the gist of the instruction even if every other word comes from a different language. SD has a lot of user models already that were trained entirely on Chinese keywords, but that was still on top of vanilla 1.5, so you can do the same kind of thing when prompting them
|
|
#
?
Apr 19, 2023 15:14
|
|
|
On MJ I asked for a picture of 一只狗在舒适的壁炉旁睡着了 which is Chinese for a dog sleeping next to a cozy fireplace. You be the judge of the results:  The German results: 
hydroceramics fucked around with this message at 15:29 on Apr 19, 2023 |
|
#
?
Apr 19, 2023 15:25
|
|
|
hydroceramics posted:On MJ I asked for a picture of 一只狗在舒适的壁炉旁睡着了 which is Chinese for a dog sleeping next to a cozy fireplace. You be the judge of the results: I just entered the Japanese for dog and got just dogs. I'm not saying it understands Chinese language structure, but if you just entered the Chinese for dog there's probably photos in the data tagged dog in Chinese.
|
|
#
?
Apr 19, 2023 15:27
|
|
|
|
| # ? May 30, 2024 19:02 |
|
|
hydroceramics posted:On MJ I asked for a picture of 一只狗在舒适的壁炉旁睡着了 which is Chinese for a dog sleeping next to a cozy fireplace. You be the judge of the results: they're gonna ban you for weed scrolling I thought it was a day early! oohhhh I do need to make Stable Diffusion bongs tomorrow don't I?
|
|
#
?
Apr 19, 2023 15:28
|
|
