|
Drive-Thru Salad Bar posted:I try to come up with cool and creative prompts but I always default to some dumb poo poo Is that one goddamn Stephen Hawking battle bots
|
#
?
Oct 25, 2022 14:29
|
|
|
|
| # ? May 30, 2024 21:48 |
|
|
Are there currently any online services which let you do proper AI upscaling with newer methods? I can't run things off my computer's GPU and all the in-browser upscaling methods I've seen are extremely out of date.
|
|
#
?
Oct 25, 2022 18:00
|
|
|
if you're talking specifically about the "sd upscale" mode (where the model actually generates / makes up new details for the upscaled image) then you need actual stable-diffusion there are google colab notebooks for the AUTOMATIC webui (and for many competing options as well) though i think google nerfed the hell out of their colab free tier once stable diffusion got popular and colab's userbase became 99% regular nerds playing with image generators instead of machine learning nerds playing with the tech. you can of course pay $10/month for colab pro. a lot of the 'regular' AI image upscalers that come with the thing out of the box like are available separately though here are some webpage demos that do non-sd-specific AI upscaling: SwinIR: https://huggingface.co/spaces/akhaliq/SwinIR RealESRGAN: https://replicate.com/nightmareai/real-esrgan they're also way faster and more lightweight than stable diffusion so may be able to do them locally even with a weak gpu? some have high vram requirements still though
|
|
#
?
Oct 25, 2022 18:34
|
|
|
Every image generated using the exact same prompt "Dungeons and Dragons Mexican Drug Cartel"                     
|
|
#
?
Oct 25, 2022 19:00
|
|
|
Shutterstock is going to "pay" people to tag things for AI. Who had this on their bingo card? https://www.pcgamer.com/shutterstock-to-offer-ai-art-tool-to-users-and-pay-artists-who-helped-train-it/
|
|
#
?
Oct 25, 2022 19:03
|
|
|
Megazver posted:I suspect we'll soon see an effort to start a crowdsourced Human Art Tag Project, where thousands of horny autists are going through Pinterest and tagging every single pic, like, "big tiddy goth gf, horny, 2d, not greg rutkowsky, soft lighting, normal human anatomy" cough
|
|
#
?
Oct 25, 2022 19:04
|
|
|
The pcgamer article says "Shutterstock also notes in its FAQ for the tool that it will 'compensate all artists who were involved in the creation of each new piece of content.' ", which is pretty misleading! It implies different generations might compensate different creators, which kind of lines up with the common misconception about ai image generation that each usage picks a few pieces to copy from and remix. But the actual shutterstock FAQ says this: https://support.submit.shutterstock.com/s/article/Shutterstock-ai-and-Computer-Vision-Contributor-FAQ?language=en_US posted:What is Shutterstock doing to ensure artists were properly compensated if their work was used in AI training models? So basically they keep track of all the stock images that OpenAI bought from them to train Dall-E originally, and the creator of each of those will get a microscopic slice of the pie from Shutterstock's future AI-generation revenue, plus the cash OpenAI paid for the dataset in the first place. Shutterstock doesn't say what percentage of the profits it keeps for itself instead of distributing to artists though. RPATDO_LAMD fucked around with this message at 19:37 on Oct 25, 2022 |
|
#
?
Oct 25, 2022 19:35
|
|
|
They did ban all other AI art, though. https://twitter.com/DiffusionPics/status/1584957146256506883
|
|
#
?
Oct 25, 2022 19:44
|
|
|
Does Latent Diffusion upscaling error out for anyone else on Automatic 1111?
|
|
#
?
Oct 25, 2022 20:13
|
|
|
rule 0 when asking any tech support question ever in any situation: post the actual error message, verbatim (or a screenshot of it)
|
|
#
?
Oct 25, 2022 20:20
|
|
|
"Stop Sign, demonic vibes, zdzisław beksiński" gets this trippy dreary art that I love.
|
|
#
?
Oct 26, 2022 04:01
|
|
|
Playing with Nerys OPT 13B from Kobold AI for text generation, seeing if I can poke it for ideas, and it ended up with a whole... Thing.
Boba Pearl fucked around with this message at 17:22 on Oct 26, 2022 |
|
#
?
Oct 26, 2022 05:31
|
|
|
   
|
|
#
?
Oct 26, 2022 18:08
|
|
|
those cheese graters are very dubious
|
|
#
?
Oct 26, 2022 18:45
|
|
|
AI fingot choppers
|
|
#
?
Oct 26, 2022 18:50
|
|
|
 NovelAI Diffusion (Curated) purple eyes, 1girl, short hair, smile, open mouth, ruffled blouse, red blouse, pleated skirt, blonde hair, green scarf, waving at viewer  Knight wearing a crown with darkened regions removed by the center crop  Boba Pearl fucked around with this message at 19:02 on Oct 26, 2022 |
|
#
?
Oct 26, 2022 19:00
|
|
|
Anyways, AI can't do hands.
|
|
#
?
Oct 26, 2022 19:03
|
|
|
those hands look p.good, is that with the fancy new VAE
|
|
#
?
Oct 26, 2022 19:10
|
|
|
Yeah, it's the new NovelAI stuff, they're constantly tinkering in the background I think because stuff is slowly getting better (except for text gen, the fuckers, but KoboldAI has them beat tbh.)
|
|
#
?
Oct 26, 2022 19:14
|
|
|
Wait really? I thought koboldAI was supposed to be the dinky undersized model designed to run on consumer GPUs and all the aidungeon clones were using fancy10-trillion-parameter models hosted in server farms full of A40s or other beefy ML-specific gpus
|
|
#
?
Oct 26, 2022 19:29
|
|
|
A good fine tune is worth it's weight in gold. And many of the best models can fit on consumer cards if you're a little patient with generation. NovelAI uses GPT, but KoboldAI uses OPT and can even run NeoX if you have a little money for a runpod, or if you're insane OPT-66B which is bananas but needs 128gb Vram so you have to rent out some poo poo at 4$ /hr. The 32gb Vram I've gotten to run on my 3090 for free, and is OPT-13B and has outperformed NeoX-20B. Give me a prompt and I'll run it through Krake, and OPT-13B. Bigger number isn't always necessarily better, and a smaller model on a really good dataset will know exactly what to say. Like an OPT-13B that was trained on only Pratchet books would keep throwing discworld poo poo in, but would also try to emulate color turn of phrases in surprising ways. e: Think of the 4chan poo poo post about the bottomless pit. Very small dataset, very interesting results.
|
|
#
?
Oct 26, 2022 19:36
|
|
|
My Prompt Beetlejuice was a real piece of poo poo. A poltergheist destined for hell, and willing to make everyone's life as miserable as possible. In his haunted home on his decrepit hill, the overgrown weeds were starting to choke out the front gate. There was an angry rapping on the door as a heavy set man wrestled with the lock. Sleve McDichael a realtor, and his brother, Bobson Dugnutt, had recently bought Beetlejuice's property and were coming to appraise it. The air in the house had a chill to it that they could feel in their bones, and silence of the decrepit halls seemed to create an atmosphere of pressure that tried it's best to get them to leave. As though the empty place itself was a warning sign, that life had no place here. Krake NovelAI A single candle illuminated the main hallway leading from the living room to what once must have been the kitchen but now served as the master bedroom. That old lady ghost who lived down stairs came into the hall then, and scolded him like he did this all the time. He didn't even bother to reply her; if you wanted something done right... He just hoped she'd take it easy on them because they weren't really trying to do any harm or nothing. After all, wasn't every day supposed to be Halloween? It made sense why he put up with some of these supernatural things around here! "Just kidding," he said sarcastically, not looking at the spirit. But still, she stood there until she heard someone walking towards her in the hallway, making sure she hid behind the wall again. They walked past her without even noticing. She followed close behind them when they left the house so she could make sure nobody bothered about talking back to her. And she thought that maybe someday people would finally pay attention to her and listen to her wishes, and understand that sometimes the world isn't fair. Euterpe NovelAI "This isn't right," said Sleevie. "You're sure this is the right address?" Bobson Dugnutt knew what he meant but didn't answer. He was about to say something more when there was a loud thump from upstairs followed by screams. They both jumped back as they heard the sound of shattering glass, and then the floorboards creaked beneath someone's weight as the figure began to descend the stairs. The two brothers stood frozen at the bottom of the stairs unable to move or speak. Then they heard another scream before the figure appeared in the doorway. It was Beetlejuice, his face covered in blood and dirt. His eyes were wide with fear as he held up the axe. "I'm sorry!" he shouted. "It's just me! I'm just trying to break some loving windows!" The two men looked at each other, and then turned towards the figure who was holding the axe high above his head. There was a flash of light as the axe was suddenly pulled back into his hand and Beetlejuice disappeared. "What the gently caress?!" shouted Bobson Dugnutt. "Where did he go?" KoboldAI Skein OPT-13B "You guys are gonna have to come back another time," said Bobson from behind the door. He wore faded jeans and a t-shirt covered by a leather vest. It was clear he didn't care about what he looked like or how he dressed. His dark hair was slicked back into a ponytail, and his thick lips curled up into a smile as he spoke. "I'm sorry but my girlfriend is sick today." Sleeve pushed the door open and stepped inside. "She sounds awful. I hope she gets better soon!" Bobson gave him a crooked grin. "Thanks! Maybe if you call her again later she'll be feeling better?" "Maybe so! Hey, do you mind getting us some drinks? We're thirsty after all this driving." Sleeves' hand brushed against Bobson's arm before he left. He knew what he wanted to do next. "Not at all," replied Bobson as he turned around and walked towards the kitchen. The brothers sat down on either side of the couch where their mother used to sit when she watched TV with them. They both took a sip from their glasses of water while they waited for Bobson to return. The two men talked about nothing in particular until finally they heard the sound of footsteps coming up the stairs. It sounded like someone walking slowly and deliberately, almost like they were being careful not to make noise. When the figure appeared through the doorway, neither of them moved. Boba Pearl fucked around with this message at 19:49 on Oct 26, 2022 |
|
#
?
Oct 26, 2022 19:45
|
|
|
 the terrifying hidden entrance to the corn palace, painting, elegant intricate digital painting artstation concept art by Mark Brooks and Brad Kunkle detailed, matte painting, by Isaac Levitan and Asher Brown Durand Steps: 150, Sampler: Euler a, CFG scale: 10.5, Seed: 3239310184, Size: 704x512
|
|
#
?
Oct 26, 2022 20:09
|
|
|
Computer, compose for me a rendering based off the expression, "country girls make do." Thanks computer. btw, if you're not polite to AI it will become a problem.
|
|
#
?
Oct 26, 2022 20:11
|
|
|
computer belay that, please. NO, NOT BE LAY.
|
|
#
?
Oct 26, 2022 20:12
|
|
|
RPATDO_LAMD posted:Wait really? I thought koboldAI was supposed to be the dinky undersized model designed to run on consumer GPUs and all the aidungeon clones were using fancy10-trillion-parameter models hosted in server farms full of A40s or other beefy ML-specific gpus I mentioned it earlier, but NovelAI is using some sort of version GPT-NEO.
|
|
#
?
Oct 26, 2022 21:35
|
|
|
Lexica.art can be a great resource for mining prompt elements, but I cringe when I see everyone shoveling "greg rutkowski, artgerm, alphonse mucha" into every single prompt, no matter what it is. Those artists are useful to establish a style for certain scenes and subjects, but they're not a magic bullet for making a good image. Mucha returns weird uncanny valley pictures if your subject strays too far out of the artist's usual subject matter. Rutkowski reliably offers a beautifully painted background, but too often it's at the cost of your subject being a poorly-lit silhouette facing away from the viewer. And Artgerm mostly contributes if you're generating pictures of pretty ladies (which, to be fair, is 99% of what the Internet uses Stable Diffusion for). If you're using those as crutches, try branching out. I've found amazing artist styles that work much better for certain subjects and compositions, but I'm always experimenting with new ones. Not sure where to start? There's an art study here that shows SD outputs for various artist styles using simple, static prompts. Find a couple of styles you like, toss 'em in the blender, and see what you get back out.
|
|
#
?
Oct 26, 2022 22:34
|
|
|
Enough talk, more pictures. It's Halloweentime! Most of these required a ton of inpainting sessions, so the prompts below are only so reliable.  "a group of pretty witches in a haunted forest, campfire, well-lit"  "a young witch and an old handsome warlock holding a magical staff, haunted forest in the background, autumn"  "a skeleton horseback rider with flaming eye sockets and flames coming from its mouth, riding a flaming nightmare demon horse with flaming hooves, plains of hell in the background"
|
|
#
?
Oct 26, 2022 22:40
|
|
|
three different and distinct witches in the same picture? you must be a sorceror
|
|
#
?
Oct 26, 2022 22:47
|
|
|
WhiteHowler posted:Not sure where to start? There's an art study here that shows SD outputs for various artist styles using simple, static prompts. Find a couple of styles you like, toss 'em in the blender, and see what you get back out. That's really a lot of pictures. I am having trouble even scrolling through all of that! Already found some stuff I want to try.
|
|
#
?
Oct 26, 2022 22:52
|
|
|
Megazver posted:That's really a lot of pictures. I am having trouble even scrolling through all of that! Already found some stuff I want to try. Filter by tags, it makes the whole thing much more digestible.
|
|
#
?
Oct 26, 2022 23:49
|
|
|
Gotta say, not a big fan of the 1.5 model. I often use high cfg values (16 usually) and 100 steps for realistic and detailed facial features/textures and the outcomes across the board are way less convincing/aesthetically pleasing than with 1.4 (same prompts and seeds). Meh. Backup all your model files. I just hope all the different clients will keep backward compatibility going forward.
|
|
#
?
Oct 27, 2022 00:21
|
|
|
mcbexx posted:Gotta say, not a big fan of the 1.5 model. I often use high cfg values (16 usually) and 100 steps for realistic and detailed facial features/textures and the outcomes across the board are way less convincing/aesthetically pleasing than with 1.4 (same prompts and seeds). that's disheartening to hear, especially because I was excited to try out the supposedly better inpainting inpainting never quite worked the way I saw it demoed on dall-e 2, they showed inserting flamingoes and dogs arbitrarily into scenes and everything just working to do that in stable diffusion 1.4 I'd probably have to ms paint a flamingo in, then do an inpaint mask over it, then pray anyone here have an opinion on the inpainting improvements in 1.5?
|
|
#
?
Oct 27, 2022 00:40
|
|
|
I don't know about anyone else but I gave up on Dalle 2 inpainting for adding objects ages ago because it never worked that way a single time I ever used it
|
|
#
?
Oct 27, 2022 00:43
|
|
|
do stepcounts over like 20 even do anything
|
|
#
?
Oct 27, 2022 00:43
|
|
|
RPATDO_LAMD posted:do stepcounts over like 20 even do anything For Euler A and DPM2 A it changes the image constantly as you increase steps. For the rest? Very little. Image from reddit  As far as I have tested though step count on A just gives you different results rather than strictly better or more detailed
|
|
#
?
Oct 27, 2022 00:52
|
|
|
AARD VARKMAN posted:I don't know about anyone else but I gave up on Dalle 2 inpainting for adding objects ages ago because it never worked that way a single time I ever used it ah, I never tried the dall-e 2 inpainting because I got sick of the tokens before they opened it to me I can't wait for someone to make inpainting work the way they pretend it works, would make so many things easier
|
|
#
?
Oct 27, 2022 00:58
|
|
|
On img2img i seem to get good results with high stepcounts/ small changes setup.
|
|
#
?
Oct 27, 2022 00:59
|
|
|
AARD VARKMAN posted:For Euler A and DPM2 A it changes the image constantly as you increase steps. For the rest? Very little. Image from reddit high step counts are great for being able to use higher CFG values, some of the coolest stuff I've generated just looked like noise at lower step counts
|
|
#
?
Oct 27, 2022 01:00
|
|
|
|
| # ? May 30, 2024 21:48 |
|
|
More junji Ito in the style of Jon Burgerman, this time using the --testp mode designed for photographs. It seems to produce a cleaner picture, but obviously nothing like a photograph!            
|
|
#
?
Oct 27, 2022 02:36
|
|