|
Got some amazing fidelity on this one: "a baby seal that eats only Doritos and drinks only Mountain Lion on an apartment carpet floor"
|
# ? Oct 22, 2022 23:05 |
|
|
# ? May 30, 2024 22:38 |
|
Mercury_Storm posted:Got some amazing fidelity on this one: you liar! thats just a normal picture. your no AI artist
|
# ? Oct 22, 2022 23:07 |
|
NOOOoooo! how will my reputation as an AI artist recover from this gaffe!!
|
# ? Oct 22, 2022 23:16 |
|
Rutibex posted:you liar! thats just a normal picture. your no AI artist You can tell because the text is not messed up Turns out that making text is harder than making images Tamba fucked around with this message at 23:28 on Oct 22, 2022 |
# ? Oct 22, 2022 23:24 |
|
Tamba posted:You can tell because the text is not messed up also im sure "drinking mountain lion" would produce very different results
|
# ? Oct 22, 2022 23:27 |
|
Tamba posted:You can tell because the text is not messed up e: a cute plushie baby seal orb on a carpet floor, surrounded by doritos and mountain dew
|
# ? Oct 22, 2022 23:32 |
|
"a baby seal that eats only Doritos and drinks only Mountain Lion on an apartment carpet floor"
|
# ? Oct 22, 2022 23:43 |
|
plush baby seal on carpet, doritos bag, mountain lion
|
# ? Oct 22, 2022 23:56 |
|
mobby_6kl posted:SD seems to have absolutely no idea who Beavis and Butt-head are but the results are... something if you think that's bad you should see Midjourney's take daguerrotype photograph of beavis and butthead in an early earth jungle bonus beavis and butthead cuddling on the couch, by studio ghibli and alex ross
|
# ? Oct 23, 2022 01:53 |
|
I gave Lexica.art the Beavis & Butthead challenge, and it attacked me with this Beavis and butthead :: 8k resolution concept art by magali villeneuve :: patrice murciano :: minjae lee :: beautiful lighting :: hyperdetailed :: intricately detailed splash art trending on artstation :: deep colors :: unique striking composition :: james jean :: jonas de ro :: alena aenami :: sharp focus I appreciate that it book-ended the set with each of them. Attack on Princess fucked around with this message at 03:20 on Oct 23, 2022 |
# ? Oct 23, 2022 03:12 |
|
lexica.art is probably the best free SD service and the prompt search/image search tools are useful and fun. You even can adjust the image size. The outputs aren't high quality and you can't pick the sampler but they'll do the job for just messing around, you can always take the same prompts and seeds and use a different service for fine tuning. They also upscale automatically. I'm not sure what the usage limit is, it's come back with "try again later" a couple of times, then immediately gone back to spitting out pics. "Rainbow Lion, Street Art"
|
# ? Oct 23, 2022 11:43 |
|
it seems as though DALLE finds pokemon quite confusing "an Oddish and an Eevee happily walking along a woodland path, anime"
|
# ? Oct 23, 2022 17:05 |
|
"Christmas card reindeer and Santa by Anton semenov" "a man made out of the best bread ever, artisnal bread, bread man, crust" Rutibex fucked around with this message at 18:29 on Oct 23, 2022 |
# ? Oct 23, 2022 18:12 |
|
they are working on the fourth iteration of the waifu-diffusion model already this one will have 10-20 million anime jpegs in the dataset, compared to just 50k for the first version (it's still based on sd 1.4 though, not 1.5) seems like the money to make bigger and bigger finetunes is rolling in fast
|
# ? Oct 23, 2022 19:48 |
|
BrainDance posted:One thing I noticed is that I'm getting very different results now, less these long essays and more scenarios which is cool but I wish I understood exactly why so I could fine tine things even more. You could try tweaking the repetition_penalty, temperature and TOP_P values. The higher the repetition_penalty is, the less likely the model is to generate a token it has already seen, so the generations become more unique. Higher temperature means the model is more likely to generate less likely tokens, which is roughly how """creative""" the model becomes. TOP_P sets a threshold for the quality/likelihood of the generated tokens. Set the temperature high and TOP_P low and you get random nonsense. Set TOP_P high and it filters out most of the worst garbage. Set temperature low and the generated text takes less liberties with the prompt.
|
# ? Oct 23, 2022 20:19 |
|
|
# ? Oct 23, 2022 21:52 |
|
this are really cool
|
# ? Oct 23, 2022 22:28 |
|
LifeSunDeath posted:this are really cool i was reluctant to post the prompt because its kind of embarrassing "portrait of a mother giving milk to an alien". gives a lot of very bizarre and creepy results!
|
# ? Oct 23, 2022 23:26 |
|
I love this so much
|
# ? Oct 23, 2022 23:51 |
|
A couple of days ago I was having trouble getting to bed, thinking about the imminent collapse of our society and trying to be ever so slightly less fatalist and cynical about the future. I thought back to how close man was to the brink of extinction in the ice age, and decided to use Dreamily to generate the wording for a story I had the shape of in my head. It was surprisingly easy to direct the AI to mostly stick with what I had in mind, and definitely improve it in areas. I got inspiration from what it wrote, and used it to spark new ideas and move the story and emotions forward in ways I wouldn't have otherwise. It got off course plenty, and I had to write a few really prescriptive things in order to point it back in the right direction (mostly in the story the young man tells), but I really like how it came out, which was more or less how I had imagined. The prompt involved two cavemen walking through a snowy mountain pass.quote:"Do you remember the stories?" The young hunter whispers to his partner, who seems lost in thought as he stares into the cold, icy nothingness. The tense changes at one point, and there's probably a few minor goofs here and there, but I was impressed. I ran it through an automatic plagiarism filter and it picked up a decent number of sentences yanked from everything from random nonfiction YouTube videos to old half-forgotten pulp fiction. Once these tools reach full maturity and the copyright issues are cleared up, it's really going to be scary how fully they'll be embraced by authors. I like this story enough that I decided to adapt it into a comic book as an experiment: This also involved a ton of image generation for every panel in MidJourney, picking the handful that felt close to what I had imagined or pleasantly surprised me, then selectively erasing things and adding in new prompts in DALL-E and expanding the scenes. I also manually adjusted the colors here and there to match a bit more, but nothing is painted or added by hand. After a certain point, what I was able to do with the tool started to make the visuals feel very repetitive, so I stopped. I couldn't get it to do specific poses or compositions which would add a ton to the visual storytelling, and it looses a lot of emotion that's present in the prose because of it. I couldn't even get them to consistently hold spears or bows, or show emotions easily. Still, a fun experiment. feedmyleg fucked around with this message at 04:05 on Oct 24, 2022 |
# ? Oct 24, 2022 03:52 |
|
Midjourney anime beta is looking sharp
|
# ? Oct 24, 2022 05:26 |
|
This post inspired by a couple of iconic photographs. https://news.mit.edu/2016/scene-at-mit-margaret-hamilton-apollo-code-0817 https://www.pbs.org/newshour/science/katie-bouman-hardly-knew-what-a-black-hole-was-her-algorithm-helped-us-see-one It is simply amazing how far we have come in so short a time. I want to acknowledge the hard work that came before that let me make the next images on a home computer with just a few words. This definitely feels like cheating lol. Imgur Gallery, 40 images - Casually Peeking Into Black Holes
|
# ? Oct 24, 2022 07:00 |
|
|
# ? Oct 24, 2022 07:09 |
|
Tried to do my first thing with dreambooth, training it on some pictures of myself So, from this repo, https://github.com/XavierXiao/Dreambooth-Stable-Diffusion it just stopped at this point quote:Average Epoch time: 864.25 seconds No error message, nothing, just stopped. And it gave me a model that seems to work? It's super uncanny because I just grabbed a handful of random pictures of myself but, Is that right, like is there a sensible reason it just stopped at 20% or is this wrong and somethings broken?
|
# ? Oct 24, 2022 09:35 |
|
feedmyleg posted:A couple of days ago I was having trouble getting to bed, thinking about the imminent collapse of our society and trying to be ever so slightly less fatalist and cynical about the future. I thought back to how close man was to the brink of extinction in the ice age, and decided to use Dreamily to generate the wording for a story I had the shape of in my head. It was surprisingly easy to direct the AI to mostly stick with what I had in mind, and definitely improve it in areas. I got inspiration from what it wrote, and used it to spark new ideas and move the story and emotions forward in ways I wouldn't have otherwise. It got off course plenty, and I had to write a few really prescriptive things in order to point it back in the right direction (mostly in the story the young man tells), but I really like how it came out, which was more or less how I had imagined. The prompt involved two cavemen walking through a snowy mountain pass. This is awesome thank you! I am also currently working on a project that uses the AIs for writting and art assistance. It's gonna be a D&D adventure and hopefully it will look fancy with all of my new AI art, unlike my last two which used public domain clip art.
|
# ? Oct 24, 2022 10:17 |
Is this you btw?Rutibex posted:the best outcome Ukraninan civilians is for Russia to purge out every last western influences from the country and end the civil war. going back to the status quo of the last 7 years is not ideal Rutibex posted:If you want to argue that the lines on the map as conceived by western powers are sacred than I guess russia is pure evil. but to me this is a civil war Rutibex posted:russia isn't doing any genocide, unless you count nazis as a culture. it could be argued that Ukraine was doing a genocide against Russian speaking Ukrainians though. do you care about them? Rutibex posted:russia could just nuke the entire ukranian army and send in guys with radiation suits to occupy the rubble. i hope Ukrainians understand this, they actually can't win no matter what. winning just means escalation (USER WAS PUT ON PROBATION FOR THIS POST)
|
|
# ? Oct 24, 2022 13:29 |
I mean this very sincerely and very seriously: Do not import drama from other forums, especially in a thread dedicated to AI Art that's been chugging along fine so far. GBS is one of the few forums on this site where anyone is allowed to post (except getfiscal), and as such we don't forumban anyone. If someone's posting along just fine in a thread, leave them be. If you think they're breaking a rule, report them. If you have problems with other forums, take it up with the admins or their mods. Don't do whatever it was above. And I don't want to turn this into a discussion thread on this particular post so let's move on. If you have questions about this feel free to PM me and I'll get back to you as soon as I'm able.
|
|
# ? Oct 24, 2022 14:50 |
|
Yeah well they imported the drama to GBS themselves with horrible takes. You can make your own! "professional oil painting of a man punching a nazi_ss_officer with fist, amazing art, trending on artstation"
|
# ? Oct 24, 2022 15:59 |
|
So a really interesting way to use AI in digital art that's not being really talked about is upscaling. Automatic-1111 has the option to use SwanIR tiles, and you can use that to take a reference image and make it huge. One of the problems I run into a lot (because I photobash 3D models and trace from reference,) is that I don't always have a high quality reference at the size I need. NSFW pic, but basically, I was struggling to figure out how I'd draw this part of a scene: because it's just a small part of the canvas, I know it doesn't look small but when you're tracing you need lots of space for clean lines, and part of a much larger scene (which I can't really post because anatomic models are primarily nude, if barbie doll nude) A normal part of my process now is using this tile based upscaling to get very clean references at absurd resolutions, which also lets me get more texture in the line work, and makes the piece look more cohesive when it's down-scaled. Some webtoons are done by teams, and so have seperate people sketch, ink, and color. I was thinking how more traditional artists, until the AI gets to the necessary points, could use it to massively up scale the sketches as they pass them from the sketcher to the inker, and how that might help get smoother and cleaner lines.
|
# ? Oct 24, 2022 16:24 |
|
The prompt was "puppy races". I'm not sure what foul sorcery brought this about. I did get a better one:
|
# ? Oct 24, 2022 17:25 |
|
Not sure if this is even the right place to ask, but what exactly is the “best” ai writing tool out there right now? I was futzing around with Novel AI yesterday and it seems to have a lot of neat features to work with.
|
# ? Oct 24, 2022 21:02 |
|
It's interesting that the most critical and expensive part of making a good ai model is a properly tagged dataset of source images to train on. Which the image stock companies posess. But, the existence of such an ai model directly threatens their own existence. So I'm wondering what the conversations are like in the boardrooms of places like Shutterstock. Do they attempt to create their own ai or do they wait for their own inevitable demise?
|
# ? Oct 24, 2022 21:27 |
|
MechanicalTomPetty posted:Not sure if this is even the right place to ask, but what exactly is the “best” ai writing tool out there right now? I was futzing around with Novel AI yesterday and it seems to have a lot of neat features to work with. Unless you need it tailored for a specific purpose like copywriting, go directly to GPT3 through openAI: https://openai.com/api/ https://beta.openai.com/playground Their best model (Davinci) ends up costing about $0.25 per 1k words. Their cheaper models like Ada are like, 10k+ words per $0.01 - and you start off with $18 in free credit. You do have to give them your phone number though. Definitely recommend reading the manual if you want to get serious about it, but it can do an enormous variety of things you wouldn't expect out of a text AI. You could probably subcontract like 80% of your average computer toucher office job to it - it can write your emails, it can bugfix your code, it can write your pivot tables, it can translate natural language to code... AFAIK there's no other writing AI that really competes, other than purpose-built ones for specific tasks, and those may or may not be based on gpt3 in the first place. I am less familiar with text AI than image AI though e: FWIW I just looked at NovelAI and it is as far as I can tell based on GPT1 so it has cool features but they'd be pretty far behind GPT3 on actual power (this is not at all a 100% accurate comparison but imagine GPT1 as someone who is conversational in English but doesn't grasp its nuances, vs GPT3 as a fluent and somewhat eloquent English speaker), which is unfortunately the case with most GPT-based tools I believe. deep dish peat moss fucked around with this message at 23:56 on Oct 24, 2022 |
# ? Oct 24, 2022 23:20 |
|
deep dish peat moss posted:Unless you need it tailored for a specific purpose like copywriting, go directly to GPT3 through openAI: Based on the leak It looks like it uses GPT-Neo
|
# ? Oct 25, 2022 00:00 |
|
KoboldAI is free but runs on your GPU so if you have a top of the line gaming PC that could be an option for you.
|
# ? Oct 25, 2022 01:35 |
Boba Pearl posted:KoboldAI is free but runs on your GPU so if you have a top of the line gaming PC that could be an option for you. Most model options are pretty bad though, unless you have a ton of VRAM and system RAM (for example the 20B models can run on 10GB VRAM but you need 32GB of RAM and it takes minutes to generate even a bit of text)
|
|
# ? Oct 25, 2022 02:11 |
|
Yeah I have a 3090, but I run everything on runpod which is about 1$ an hour. I'm looking at Jasper but it's expensive, and I don't want to pay a sub AND per word, which means NovelAI but that's pretty trash because it's not a very good Neo tune, and whatever they use loses a lot of info even with world information.
|
# ? Oct 25, 2022 02:32 |
|
Messing with clips from 2001: A Space Odyssey https://i.imgur.com/C1zwzIy.mp4 The furry_tranformation hypernetwork that came with the novelai leak can lead to some cursed faces. https://i.imgur.com/XkstV6c.mp4 I like how it keeps his stare mostly intact even with the denoise turned up to 0.5, and how it can't decide how many mouths he should have
|
# ? Oct 25, 2022 08:14 |
|
I try to come up with cool and creative prompts but I always default to some dumb poo poo
|
# ? Oct 25, 2022 09:53 |
|
|
# ? May 30, 2024 22:38 |
|
Drive-Thru Salad Bar posted:I try to come up with cool and creative prompts but I always default to some dumb poo poo hot drat lol
|
# ? Oct 25, 2022 14:24 |