>>29925 how does it feel knowing eurobeat has had less cultural impact than dubstep(or by extension, trap)? Which is even more funny when you consider that eurobeat is the trash pop genre of electronic music. It's not deep or complex - it's just discount rave.

>>29927 >how does it feel knowing that zoomio-normogroids don't care about the thing you love? Feels genuinely great, thank God.

>>29854 by the way, nightshade "attacks" both diffusion and classifier models (i.e. wd tagger) it accomplishes that by taking the image, taking a certain noise timestep, adding then removing the noise with "dirty label" prompt and then mixing this shit back into the original image it's as effective as shit ass tagging, especially if you choose "dirty label" to be completely random, if you choose to bleed infrequent concepts into frequent but similar ones, you'll have a much stronger effect. basically it relies on so-called "tag bleeding" nightshade targets niche concepts, so if you have 1M images to train and your target concept is just 0.001% of that (so 10 images), you only need say 2-3 images to completely throw off this concept, if you were unlucky enough to land enough epochs on tampered images, on the tampered timesteps HOWEVER popular concepts are significantly harder to fuck with, if your target concept has at least 50 images, good luck trying to throw it off lol

>>29929 so it works like normal tag poisoning, got it not something to worry about then except for e621 nerds lol

>>29929 >basically it relies on so-called "tag bleeding" So basically it's gonna be next to useless on anything but CLIP? What's stopping me from giving it the lightest denoising pass with the shittiest denoising model I can find and completely bypassing their garbage?

>>29931 If a model uses a different text encoder and VAE than the ones they target, I think it should be completely immune, other than the image looking a bit noisier. And I think they targeted SD1.5?

>>29930 ironically you can try to fight it by using seemingly false positives on tagger models, these tags aren't necessarily "false", they may be detecting shit prompts that got encoded into the image by nightshade >>29931 tag bleeding is an achilles heel of any model that uses tokenizer (all llms) but regardless, it's actually more like "semantic tag bleeding" so you can actually bleed "vanilla_(nekopara)" into "cat_girl" however nightshade specifically relies on clip to provide false guidance so you'll be effectively distilling your model off of clip >What's stopping me from giving it the lightest denoising pass with the shittiest denoising model I can find and completely bypassing their garbage? you don't have to bother to do that, this shit will not work even if you sit doing absolutely nothing there's a good chance newer models like 4o that don't rely on diffusion aren't affected in the first place, it'll just have the same effect as garbage captions would have

>>29933 >you don't have to bother to do that, this shit will not work even if you sit doing absolutely nothing So hold on, they created a poisoning method that does absolutely nothing in the end and they're jerking each other off over it anyway?

>>29934 That's all those anti-AI techniques in a nutshell. It's just a second level grift.

>>29935 funny they're even more behind than local development

>>29934 it does work under very specific conditions and these conditions are generally unachievable unless you decide to train your model sd 1.5 way on a shitton of images poisoned in the same exact way >they're jerking each other off over it if you're talking about artfags, absolutely yes i don't think there's a serious discussion about it, even a relatively minor paper such as minsnr got twice as much citations

>>29927 >had less cultural impact What cultural impact does your life have, 8non?

>>29926 you're part of the problem

>>29842 Have you tried plantmilk?

https://rule34.xxx/index.php?page=icame grim

>>29940 plantmilk oat had a serious artifacting issue unless you used karras scheduler, so I didn't even touch the later versions oh but nta

>>29941 >cartoons had more sexual impact than anime

for me it's honkers.

>>29941 Yes hentai is unpopular weeaboo garbage and not even the average gooner wants to look at it.

>>29946 then explain why Princess Peach is so high in the rank

>>29947 nintenddogshit is clearly ACO in nature

nintendo is a japanese company thoughbeit

>>29947 Almost all Peach art leans more towards cartoony rather than anime. And that's a good thing.

Then everything changed when the Irony Nation attacked.

>>29941 cob won

the fuck is a cobson?

>>29953 like dobson but weaker

>reforgeui >[DISCONTINUED] what?

>>29953 wojak variation

oni pussy

>>29957 thats a childs

>>29958 Is this childs in the room with us right now?

Stop shitposting.

>>29957 eh, sure