|
Tears... https://kotaku.com/deviantart-ai-art-nft-crypto-scam-hustle-protest-stable-1849778362 I haven't read something so close minded in a while. So demeaning like the "people who think they're artists" line and trying that old sly trick of association by likening ai artists to NFT pushing crypto tech bro's to create distaste and aversion
|
#
?
Nov 14, 2022 03:40
|
|
|
|
| # ? May 31, 2024 00:03 |
|
|
quote:That thing was, of course, AI-generated art, something that sounds cool in theory but which in practice has been little else but a chance for tech goons and grifters to steal art, kid themselves into thinking they’re artists and/or deprive working artists of work. they got our number 
|
|
#
?
Nov 14, 2022 03:44
|
|
|
Rutibex, are you using stable diffusion for these images or something else? Nothing I'm getting looks nearly as good as yours. ninjoatse.cx fucked around with this message at 04:35 on Nov 14, 2022 |
|
#
?
Nov 14, 2022 04:24
|
|
|
Rutibex has been using midjourney afaik. And yeah it seems like it is way better at turning even simple or abstract prompts into coherent artsy-looking images. For a your question on offline stable diffusion samplers, there are basically only three things that matter:
The extra randomness is all from the 'ancestral' samplers whose names end in _a, like euler a, DPM++2S a, etc. Basically the _a samplers are 'better' than the normal samplers at lower step counts but in exchange they are less stable -- generating an image with euler_a at 12 steps, 24, and 36 will produce really different images. So if you just want to generate one image and be done with it they are excellent but if you want the kind of workflow where you generate a bunch at lower stepcounts and then say "hmm i want a more refined version of this one" and rerun it at a higher count, you should avoid them. For #2 and #3, they are pretty closely tied together. A lot of the samplers like DPM2 are based on mathematical work to come up with a cool formula that converges in 'fewer steps' but aren't actually more practically useful, because they take twice as long to run but don't quite halve the number of steps needed for a good image. So pretty much you only really want to care about one ancestral sampler and one non-ancestral sampler, whichever ones are 'fastest' at producing an image. DPM++ is afaik the newest and most cutting-edge one and basically the best, with a fast runtime with comparable speed to euler while also converging fast in usually ~20 steps although up to 25-30 with some prompts. Personally I use nothing but euler a at ~12-20 steps for the best speed/quality tradeoff, DPM++ 2M Karras at ~15-25 steps for a non-ancestral sampler. (I haven't really compared euler_a with DPM++ 2S a (the ancestral version of the dpm++ sampler) yet, but I got the feeling that it was worse than euler at sub-20 stepcounts which are the ones I want to use it at.) If you're using an older version of the webui that doesn't have the new DPM++ samplers, just use euler or ddim at 20-30 steps, sometimes bumping up to 50-100. They are the ones with the fastest runtime, and in terms of image quality euler at 40 steps is gonna beat heun or dpm2 at 20 steps any day while taking the same time to generate. RPATDO_LAMD fucked around with this message at 04:57 on Nov 14, 2022 |
|
#
?
Nov 14, 2022 04:53
|
|
|
RPATDO_LAMD posted:*snip* Thanks for the write up. I found some online sources and looks like there's a bit more homework than just plugging and chugging.
|
|
#
?
Nov 14, 2022 05:02
|
|
|
What does it mean when a sampling method is 'Karras' or 'a Karras'? I gather that's a person but  Like there's DPM2, DPM2 a, DPM2 Karras, and DPM2 a Karras. Is there meaning in those terms? I like DPM++ 2M at 25 steps. vvv awesome, thank you! Mozi fucked around with this message at 05:59 on Nov 14, 2022 |
|
#
?
Nov 14, 2022 05:03
|
|
|
Karras is named after this paper by Tero Karras. I won't pretend I understand the math involved but basically it proposes a way to improve samplers by adding some random noise to the time step on the sampler's equation-solving process itself to make it less deterministic. (Samplers work by solving differential equations btw... I'm not exactly sure how that works.) It's supposed to lead to better-quality images at lower step counts. The paper also covers the topic of "hey, all y'all other researchers have been treating these diffusion image models as monoliths and using one specific sampler with each pre-trained model (matching the sampler it was trained with) but actually you can swap around samplers on the fly and use any sampler with any model." So we probably have Karras back in June 2022 to thank for the fact that we are swapping between different samplers in the first place. This is totally unrelated to the way ancestral samplers add noise to the image, so unlike the _a samplers the karras samplers won't constantly mutate the image. But they do produce a different image for each seed compared to the other non-ancestral samplers. So the "Karras" samplers basically use the stochastic scheduler from that karras paper to become a bit better at lower step counts. Afaict there's no reason not to use them, unless you're trying to reproduce a specific image that was made with a non-karras sampler. The "a" samplers are ancestral samplers: RPATDO_LAMD posted:Basically the _a samplers are 'better' than the normal samplers at lower step counts but in exchange they are less stable -- generating an image with euler_a at 12 steps, 24, and 36 will produce really different images. RPATDO_LAMD fucked around with this message at 05:30 on Nov 14, 2022 |
|
#
?
Nov 14, 2022 05:27
|
|
|
Rutibex posted:"optical illusion paradox" is a banger of a prompt. Its like traveling through the back rooms or something Oh, wow This is something I would have previously put on an AI Turing test. That's incredible.
|
|
#
?
Nov 14, 2022 06:33
|
|
|
|
|
#
?
Nov 14, 2022 08:32
|
|
the Mods, they knew!
|
ninjoatse.cx posted:Rutibex, are you using stable diffusion for these images or something else? Nothing I'm getting looks nearly as good as yours. Oh sorry I should have specified, it's midjourney v4
|
|
#
?
Nov 14, 2022 10:04
|
|
|
RPATDO_LAMD posted:Karras is named after this paper by Tero Karras. In my experience the Karras samplers make cleaner results at low step counts, but will sometimes start introducing weird artifacts to the image somewhere after 50 samples. I've mostly been using 50-80 for my samples range so I've seen that happen a few times. Example here on reddit. https://www.reddit.com/r/StableDiffusion/comments/y2h3mg/comparison_of_the_karras_samplers_in_auto1111/ The non-karras samplers seem a lot less capable of overcooking your picture like that, though it doesn't happen too frequently.
|
|
#
?
Nov 14, 2022 10:33
|
|
|
What's better than an optical illusion paradox? An optical illusion paradox filled with people! This seems to create bizarre dystopian spaces. It makes your mind ask what the hell is going on here?!     
|
|
#
?
Nov 14, 2022 11:15
|
|
|
all of those seem to firmly obey one direction of gravity. can you get one with multiple like the original Escher painting?
|
|
#
?
Nov 14, 2022 16:00
|
|
|
AI art game played through discord launched https://twitter.com/AconiteCo/status/1592259866932183041 https://www.bureauofmultiversalarbitration.com/ Trailer for those who don't want to click through https://www.youtube.com/watch?v=tA2PKJ2TDyQ fez_machine fucked around with this message at 23:27 on Nov 14, 2022 |
|
#
?
Nov 14, 2022 23:25
|
|
|
AARD VARKMAN posted:all of those seem to firmly obey one direction of gravity. can you get one with multiple like the original Escher painting? Sorry I could never get it to do multiple gravitys, despite many many rerolls. The AI just seems to not like the concept As a consolation prize here are some pictures where the Paradox Realm intersects with the Abyss and the Skeletons Lands. lol my D&D players are hosed                    
Rutibex fucked around with this message at 11:09 on Nov 15, 2022 |
|
#
?
Nov 15, 2022 00:14
|
|
|
You did it, you generated Scarytown.
|
|
#
?
Nov 15, 2022 00:23
|
|
|
I love this
|
|
#
?
Nov 15, 2022 03:17
|
|
|
Can someone make a line of jesus candles with something funny on them.
|
|
#
?
Nov 15, 2022 04:38
|
|
|
LifeSunDeath posted:Can someone make a line of jesus candles with something funny on them. 
|
|
#
?
Nov 15, 2022 05:13
|
|
|
LOL.
|
|
#
?
Nov 15, 2022 05:15
|
|
|
I've been playing with the Mid Journey V4 Diffusion Book for Stable Diffusion. This thing was trained on a weird dataset it loves putting people into things and loves anything space related. Flying out of a deep crater in a rocket got me... a planet entering atmosphere I think.  The image quality is really good getting what you asked for on the other hand well at least img2img works pretty good for guiding it. https://huggingface.co/prompthero/midjourney-v4-diffusion using the colab link from here there are 2 instances running this right now but they both seem to crash every so often I think it's because of the multiple models and having to download them so if someone else before you ran a different one things can get messy. Don't forget to change the model to MJ the others are way rougher, Classic Disney can be fun if you are trying to get a Disney character to do something other subjects I got mostly garbled messes. e: LifeSunDeath posted:Can someone make a line of jesus candles with something funny on them.  see doesn't really get the idea of a candles with the face of jesus.
|
|
#
?
Nov 15, 2022 15:32
|
|
|
Hello AI thread, what are the current best ai tools for inpainting and outpainting with photorealistic results? A friend of mine was asking if I could correct a picture they took, using AI, by removing their reflection on a glass from a photo, even if the end result is not concordant exactly to reality.
|
|
#
?
Nov 15, 2022 17:42
|
|
|
 Dominions 4, with better graphics... Popoto posted:Hello AI thread, what are the current best ai tools for inpainting and outpainting with photorealistic results? That sounds pretty tricky - right now the best model for inpainting is the SD1.5 inpainting model (don't have a link offhand but I'm sure it's on huggingface) and it's good for inserting stuff (the above image started off as just a waterfall picture, then I inpainted the columns, the people, the elephant, etc.) But trying to inpaint some glass that needs to be transparent... that's not as simple as just adding in some new element, I'm not really sure about that. For example I just did a little test, generated an image of a man looking at his reflection in a window:  I inpainted the window and changed the prompt to 'a man looking through a window,' but my denoise wasn't high enough so I got an even better reflection:  So I changed to 'latent nothing' with full denoise and changed the prompt to 'transparent window':   Not great - as you can see the issue is that inpainting can add stuff but when you try to remove something it's really just adding something else, so you have to be able to tell it what it is you want to see other than just 'the thing now but without the reflection'. I could be missing something though! But I feel some artistic talent may be needed still. Also that's definitely his finger and thumb held up in a weird position. Definitely.
|
|
#
?
Nov 15, 2022 19:49
|
|
|
I think latent noise might be a better way to go, but in this case I would probably just use Photoshop since it's easier.
|
|
#
?
Nov 15, 2022 20:02
|
|
|
AI art is still a toy it's not yet a tool but you can see what the tool might look like.
|
|
#
?
Nov 15, 2022 20:08
|
|
|
For simple enough in-painting needs you might try lama-cleaner as a fairly friendly GUI based tool for quick edits. You can theoretically get good results from stuff like Stable Diffusion if you are willing to fiddle forever but if you just need to delete someone from a photo or fix scratches on a scan or something this usually works a treat.
|
|
#
?
Nov 15, 2022 20:16
|
|
|
the Mid Journey trained Stable Diffusion is hilarious This is from a bunch I did that were just terrible puns on Elon Musk's name I'm not positive which one this was but it was probably terrible. I can't believe how easy it is to get high quality images from this one (boy does it love doing people it gave me a paladin for "spoon"
|
|
#
?
Nov 15, 2022 20:52
|
|
|
I've been posting variations of Elon Musk as a clown until I get kicked off twitter. Here are some of them. The first few are wombo and I think the others are stable diffusion.       
|
|
#
?
Nov 15, 2022 21:17
|
|
|
Is there any web-based tool (with no complex setup) which allows me to rough in shapes and get finished images? I know the tech exists if you properly know what you're doing with this stuff, but I'm just a dummy who uses MJ on Discord. I tried loading a rough sketch of outlines into DALL-E but it just ignored it.
|
|
#
?
Nov 15, 2022 21:58
|
|
|
feedmyleg posted:Is there any web-based tool (with no complex setup) which allows me to rough in shapes and get finished images? I know the tech exists if you properly know what you're doing with this stuff, but I'm just a dummy who uses MJ on Discord. I tried loading a rough sketch of outlines into DALL-E but it just ignored it. Collage on artbreeder
|
|
#
?
Nov 15, 2022 22:01
|
|
|
feedmyleg posted:Is there any web-based tool (with no complex setup) which allows me to rough in shapes and get finished images? I know the tech exists if you properly know what you're doing with this stuff, but I'm just a dummy who uses MJ on Discord. I tried loading a rough sketch of outlines into DALL-E but it just ignored it. The Huggingface https://huggingface.co/spaces/anzorq/finetuned_diffusion is working right now, this breaks for hours at a time, but go to the img2img tab and it should do what you want, change the model from Arcane it's not very good you probably want it set to Mid Journey to start the others don't seem as well trained. That colab notebook (and the hugging face hosting it) I linked will do this with img2img it looks insane when you open it (it's code) but you just have to press run all and wait like 5 minutes get a link and click it boom website running on google colab for you that does AI art in a nice package with img2img negative prompts and of course the normal positive prompting. You can combine all of these at once. Really anything supporting img2img from stable diffusion linage should be able to do that. Just remember down your Google colab sessions so you aren't wracking up hours you haven't used and getting lower priority for being a resource hog on a free account. Google is giving you a nice GPU when your done let the next person have it (it could be someone here!)
|
|
#
?
Nov 15, 2022 22:26
|
|
|
I’m helping a friend with a project he’s doing, and trying to make some simple, stylized cutesy woodland creatures.I think I could use some prompt tips if anyone has ideas for me. I’m using Huggingface Stable Diffusion. He’s going to redraw them, so it’s the general animal design that’s important, not the technique. The background doesn’t matter, plain is best. So far I’m using variations on “2d game concept art of cute bunny woodland animal, artstation, art by Mila useche, ghibli” Mila Useche is some random artstation artist, I don’t know if it helps or not. I’m making normal images and silhouettes. Ghibli probably isn’t the best, that style might be over complicated but it’s the first thing that came to me. He liked these in particular because they’re generic cute animals he can turn into whatever he wants:  I’m not able to get that look consistently, though. Some others that were close:       Any ideas on how I can get better, more consistent results? I’m getting lots of images that are way off the mark and cherry picked the ones above. e- sorry for the stupidly big images, I was getting errors from the Awful app so I had to screenshot them 
|
|
#
?
Nov 15, 2022 22:33
|
|
|
pixaal posted:The Huggingface https://huggingface.co/spaces/anzorq/finetuned_diffusion is working right now, this breaks for hours at a time, but go to the img2img tab and it should do what you want, change the model from Arcane it's not very good you probably want it set to Mid Journey to start the others don't seem as well trained. This is perfect, thank you so much! I found some old sketches from a lost media project I want to try and generate full paintings of. My first few tests went... okay. I'll toy with it. Snowy posted:Any ideas on how I can get better, more consistent results? I’m getting lots of images that are way off the mark and cherry picked the ones above. Have you tried "animation model sheets"? This is what "animation model sheets of a cute hamster, multiple angles, chibi, in the style of ghibli" got me on MJ, first try: 
|
|
#
?
Nov 15, 2022 23:20
|
|
|
pixaal posted:see doesn't really get the idea of a candles with the face of jesus.  A few random things I've learned: Adding x-punk style makes everything better. Including punkpunk. Crystalpunk and Solarpunk can sometimes be amazing as well.         Bonus, retropunk jesus  Seed Info Imgur Album You can add "2d" and/or "3d" to any prompt or negative prompt to push it more in that direction, this works on any of the models. This is incredibly powerful if you push the weights too far, ((((((3d)))))) will generate a mesh image. This is big.   Seed Info Imgur Album I have no way to consistently get side views of the same subject and attempts can wildly alter the base image so it's unusable as of right now, but TXT2MESH will exist one day, and this seems a likely avenue to look into in making that a reality. feedmyleg posted:Is there any web-based tool https://huggingface.co/spaces/huggingface-projects/diffuse-the-rest
|
|
#
?
Nov 16, 2022 00:07
|
|
|
feedmyleg posted:Have you tried "animation model sheets"? This is what "animation model sheets of a cute hamster, multiple angles, chibi, in the style of ghibli" got me on MJ, first try: Oooooh no, thank you I’ll play with that!! 
|
|
#
?
Nov 16, 2022 00:12
|
|
|
Post what you get, I'm curious what ends up resonating!
|
|
#
?
Nov 16, 2022 00:16
|
|
|
feedmyleg posted:This is perfect, thank you so much! I found some old sketches from a lost media project I want to try and generate full paintings of. My first few tests went... okay. I'll toy with it. I'll take "what is the hampster dance?" for 500, alex
|
|
#
?
Nov 16, 2022 00:17
|
|
|
feedmyleg posted:"animation model sheets of a cute hamster, multiple angles, chibi, in the style of ghibli" That prompt is a goldmine      I took one of the leprechauns and built an image around it with inpainting (locally in SD) 
Mozi fucked around with this message at 00:20 on Nov 16, 2022 |
|
#
?
Nov 16, 2022 00:18
|
|
|
KwegiboHB posted:I don't know what you mean lol dank thanks for trying
|
|
#
?
Nov 16, 2022 04:18
|
|
|
|
| # ? May 31, 2024 00:03 |
|
|
 https://www.youtube.com/watch?v=Qto8R9j6VaI So one of the main SD youtubers has made a vid specifically pointing out how people can insert malicious code into SD models and gently caress your poo poo up. Walks through the whole situation and where the gap is, ways to protect yourself, and yeah, just all around an important step.
|
|
#
?
Nov 16, 2022 05:08
|
|
