Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
Next one was me trying to have Snoopy fight Goku DBZ style.
I think for most scenes involving two distinct characters especially if they're not in direct physical contact, you'll have a far easier time inpainting them separately. Getting it to work in a single prompt takes a pretty damn lucky seed and some clever prompting.

Tom's fucked up their shit and changed it. Here are their new charts, properly putting Nvidia's latest at the top. But with at least one mistake on them, iterations/sec on one where it should say FP16 TFLOPS.
RX7900 XT(X) looking impressive, especially considering relative price and availability. Do you know why they're not showing up in the fourth graph though?
 
My assertion is that it doesn't matter if the murder weapon sitting on your desk is a copy. The evidence is fake, but if nobody can tell or denying it makes someone's job harder, it implicates you.
But it does matter if the law you're being prosecuted under only prohibits something that is real.
 
  • Agree
Reactions: A Hot Potato
It saddens me that these benchmarks never include cards like the p40 or the m40 from Nvidia. With a little bit of rigging to add an active cooling system you can have something with a little less performance than a 3060 with twice as much vram. For around $280 on the secondary market.

I think for most scenes involving two distinct characters especially if they're not in direct physical contact, you'll have a far easier time inpainting them separately. Getting it to work in a single prompt takes a pretty damn lucky seed and some clever prompting.

I actually think duo or multi shots are the only pieces of safe commission work for artists. Yes you can inpaint, but if you want a more fluid scene like two characters dancing while embracing, you are going to have a bad time.
 
Last edited by a moderator:
That's why you don't want shit like that in front of a jury. And why even a judge might be so disgusted by it as to get the law wrong. However, something that is literally indistinguishable from CP is likely to be unprotected by the First Amendment if it has no artistic merit (however arbitrary that phrase is) and exists solely to excite prurient interest.
Even more than regular CSAM, it would be basically impossible to prosecute babyrapers if they were capable of behaving vaguely like human beings while jacking it to computer generated child porn. Exercising a modicum of shame and self-preservation. As bad as the pedo problem is, I'll bet there's a contigent who are already basically capable of just quietly downloading nasty shit securely and jacking off to it without ever exposing themselves to the world. Of course, none of those types would ever join groups like Prostasia or Boychat or whatever, because they would, one assumes, be capable of realizing that they are not, in fact, normal, and that their activity should never be normalized.

But you know it'll never happen for the pedos who are already getting caught. Will a babyraper be able to keep to just manipulating a model trained on little girls in swimwear to generate nude images? Oh no he won't. He'll try and excuse his collection of ever more thousands of images of naked kids as 'just used to generate a model, it's like the children from the photos never even existed'. Will the babyraper just quietly masturbate to himself in his CP dungeon? Hell no he won't, those images will be traded for material showing real child abuse or 'virtual' child abuse based on very real child abuse as soon as they're generated.

EDIT: Maybe I'm overly trusting in the rationality of the Ammurrican empire's legal system, but I don't think this should be too hard to get right. If it was 'obscene' for Paul Little to produce videos where he pissed on young women wearing 'slutty 14 year old' outfits, then the sort of stuff that will be generated as virtual child pornography sure as shit should be treated the same. The minimum that the pedos are going to produce is images of pre-pubescent girls naked, progressing to very explicit anatomical details. Certainly, such things cannot be justified as non-obscene on the basis that 'artistic' nudist shots that 'just happen' to be explicit enough that many pedos could consider them worth jacking off to, let alone family photos of unclothed children. This sort of thing is solely prurient in nature.
 
Last edited:
EDIT: Maybe I'm overly trusting in the rationality of the Ammurrican empire's legal system, but I don't think this should be too hard to get right. If it was 'obscene' for Paul Little to produce videos where he pissed on young women wearing 'slutty 14 year old' outfits, then the sort of stuff that will be generated as virtual child pornography sure as shit should be treated the same. The minimum that the pedos are going to produce is images of pre-pubescent girls naked, progressing to very explicit anatomical details. Certainly, such things cannot be justified as non-obscene on the basis that 'artistic' nudist shots that 'just happen' to be explicit enough that many pedos could consider them worth jacking off to, let alone family photos of unclothed children. This sort of thing is solely prurient in nature.
It's already forbidden to have the ultra realistic generated images under the PROTECT Act of 2003.

What pedos might be able to get away with is having the trained image model but none of the images used to train it, and none of the images it makes. That means they would need to make the images and delete them immediately before getting caught. There are technically no images in an image model, although I think that Stable Diffusion copyright infringement lawsuit is disputing that. It's a black box that spits out new images when you ask it to.

If a judge is not convinced by that argument, or Congress tweaks the law, the pedo is doomed. If they bet on MAGA SCOTUS to save them, lol.

RX7900 XT(X) looking impressive, especially considering relative price and availability. Do you know why they're not showing up in the fourth graph though?
"This final chart shows the results of our higher resolution testing. We didn't test the new AMD GPUs, as we had to use Linux on the AMD RX 6000-series cards, and apparently the RX 7000-series needs a newer Linux kernel and we couldn't get it working."
 
It's already forbidden to have the ultra realistic generated images under the PROTECT Act of 2003.
It's hard to know how susceptible this is to an overbreadth challenge, though. The one case where a conviction was upheld involved a sex offender on probation (Whorley) and the other (Handley) was a plea bargain in which the District Court judge questioned the constitutionality of the case. In another case (Dean), the Eleventh Circuit disapproved of the reasoning of Handley, but that case involved actual child pornography.

So the PROTECT Act is probably not facially unconstitutional and can't be invalidated in its entirety, but the exact scope of how it can be construed is open, largely because the feds have avoided prosecuting cases which would open up a constitutional challenge, instead generally offering lenient pleas involving not having to register as a sex offender, like Handley's case.

So far, nobody has taken up the offer to play "Who Wants to be a Registered Sex Offender" with their life.
 
So I did some more simple imaging and I noticed that the absence of negative prompts doesn't hurt the images. I used the anything V3 model and it turned out like this
snow1.jpg
snow2.jpg


One is DDIM the other Euler. Has anyone else noticed this? Or are negative prompts only useful when you have a lot of input prompts? Mine were: Best quality, detailed, girl in a snowy landscape
 
So I did some more simple imaging and I noticed that the absence of negative prompts doesn't hurt the images. I used the anything V3 model and it turned out like this
View attachment 4353827View attachment 4353829

One is DDIM the other Euler. Has anyone else noticed this? Or are negative prompts only useful when you have a lot of input prompts? Mine were: Best quality, detailed, girl in a snowy landscape
They're less necessary for better/more specialised models. If you're using one geared more for anime girls and you're asking for an anime girl you probably don't need many, and can just add specific negatives if it's doing a particular thing you don't want. People got into the habit because NAI was meant to be run with a specific set of default positive/negative terms hidden behind their interface.
Lately I've just been throwing this negative embedding in (sometimes with varying weights) and not really worrying about it.
 
Some more stuff I've done recently, apprantly anything v4 is a thing so I'm using that permanently now. Anyways please enjoy these.
The lighting is pretty kino on the geralt photo's. what'd you do to get that? splashing in a bunch of lighting techniques into the prompt or it just showed up on it's own?
 
The lighting is pretty kino on the geralt photo's. what'd you do to get that? splashing in a bunch of lighting techniques into the prompt or it just showed up on it's own?
It kinda showed up on its own, I didn't really have to do much of anything. But alot of the ones I did have decent lighting without me having to really interfere like this one.
03443-2580653797-geralt of rivia, the witcher, anime, dvd screenshot 80's anime fantasy film b...png
 
Ayo check out that new ai voicebot, the niggas are making the stanley parable guy talk about total nigger death and some guy made CWC sing that songs about killing all the gays and the faggots.
Yea. I see it on /pol/ a lot. Very nifty. one step closer for weebs to truly build their own wives out of their computer.

I also like how every popular and semi popular board has it's own ai threads now. Mods over there seriously need to make /ai/ and /aai/(sfw version.
 
Neural Network Generated MATI:

Null explain Sneed's Feed and Seed:
Null is closing KiwiFarms:
For Null, it's the McChicken:
Null is a Navy Seal:
 
Neural Network Generated MATI:

Null explain Sneed's Feed and Seed:
Null is closing KiwiFarms:
For Null, it's the McChicken:
Null is a Navy Seal:
It's really great but if you know how the person talks then it's noticeable. Shit is one winded run-on sentence.
 
Back