Discussion about A.I.

[fallenmystic]

I finally managed to find some time to test SDXL and I got mixed feelings about it.

First of all, my RTX3080 (10Gb VRAM) turned out to be no longer powerful enough to run it locally, so I rented a cloud instance as I did to test LLamas.

I was impressed by how good the base model is. While it was almost impossible to render anything realistic with the base SD 1.5 model, it was quite easy to generate pretty decent images just by prompting.

However, the lack of the ControlNet support (in SD.Next) and the inability to use more than one Loras was a deal breaker for me. So I'll just keep testing it and learn new features without trying anything serious until the issue gets resolved.

Here are some of the images I rendered during the test:

 

[Stungiskhan]

I finally managed to find some time to test SDXL and I got mixed feelings about it.

First of all, my RTX3080 (10Gb VRAM) turned out to be no longer powerful enough to run it locally, so I rented a cloud instance as I did to test LLamas.

I was impressed by how good the base model is. While it was almost impossible to render anything realistic with the base SD 1.5 model, it was quite easy to generate pretty decent images just by prompting.

However, the lack of the ControlNet support (in SD.Next) and the inability to use more than one Loras was a deal breaker for me. So I'll just keep testing it and learn new features without trying anything serious until the issue gets resolved.

Here are some of the images I rendered during the test:

View attachment 1358834 View attachment 1358835 View attachment 1358836 View attachment 1358837 View attachment 1358838 View attachment 1358839

Overally looks quite good, while there are still obvious problems with hands and fingers and some feet.

 

[fallenmystic]

Overally looks quite good, while there are still obvious problems with hands and fingers and some feet.

With SD 1.5, I used to fix those problems by editing a lineart image by hand in Krita. I cannot take the same approach now, since SDXL + SD.next doesn't yet support ControlNet.

Besides the issue with anatomic details, I have little control over what AI generates without ControlNet anyway,

Because of that, I won't try rendering anything serious until the feature is implemented and I get comfortable with the new functionalities added since 1.5.

 

[malins]

Anyway, there has for a while been this idea of 'subliminal messaging' embedded in pictures, videos, music etc.

AI turns out to be quite good at this, you only have to squint a bit to see the message

(imgur embedded, as if I used a thumbnail you'd see it immediately)

 

[fallenmystic]

Anyway, there has for a while been this idea of 'subliminal messaging' embedded in pictures, videos, music etc.

AI turns out to be quite good at this, you only have to squint a bit to see the message

In a similar vein, something that can be called "QR code art" has been quite popular among AI enthusiasts:

https://www.reddit.com/r/StableDiffusion/comments/15swdhi

https://www.reddit.com/r/StableDiffusion/comments/145fd7w

https://www.reddit.com/r/StableDiffusion/comments/14ebb6o

(All the images are hosted externally.)

I haven't tried myself, but they're supposed to work if you use a dedicated QR code app.

 

[Apostate]

Anyway, there has for a while been this idea of 'subliminal messaging' embedded in pictures, videos, music etc.

AI turns out to be quite good at this, you only have to squint a bit to see the message

(imgur embedded, as if I used a thumbnail you'd see it immediately)

I used a thumbnail, saw it immediately.

 

[malins]

In a similar vein, something that can be called "QR code art" has been quite popular among AI enthusiasts:

I haven't tried myself, but they're supposed to work if you use a dedicated QR code app.

Tried it.

Amazingly yes, the first and third one scan as a valid QR code.
Couldn't get the second one to work but I guess it'll just depend on the scanner app, I'm using something stupid, free and old ...

Anyway most "subliminal messages" in general don't really work all that well on humans... it might be interesting to see how various AI systems react to attempts to hack their pattern recognition.

Might be an idea for a story, AI guards and policemen are introduced, which are certified to be neutral, objective and free of political, racial bias etc.
It's cheaper, they never turn their bodycams off, they're incorruptible, are always proportional in their application of force, etc.
They slowly start going rogue though, because someone has managed to introduce AI-specific subliminal messaging in their surroundings that tell them the opposition and their followers are inherently disposed to crime, etc...

 

[Frank Petrexa]

Tried it.

Amazingly yes, the first and third one scan as a valid QR code.
Couldn't get the second one to work but I guess it'll just depend on the scanner app, I'm using something stupid, free and old ...

Anyway most "subliminal messages" in general don't really work all that well on humans... it might be interesting to see how various AI systems react to attempts to hack their pattern recognition.

Might be an idea for a story, AI guards and policemen are introduced, which are certified to be neutral, objective and free of political, racial bias etc.
It's cheaper, they never turn their bodycams off, they're incorruptible, are always proportional in their application of force, etc.
They slowly start going rogue though, because someone has managed to introduce AI-specific subliminal messaging in their surroundings that tell them the opposition and their followers are inherently disposed to crime, etc...

Polley Plummer wrote a story about robotics called "Rome World".
https://cruxblog.blogspot.com/2016/01/rome-world.html

 

[Stungiskhan]

Tried it.

Amazingly yes, the first and third one scan as a valid QR code.
Couldn't get the second one to work but I guess it'll just depend on the scanner app, I'm using something stupid, free and old ...

Anyway most "subliminal messages" in general don't really work all that well on humans... it might be interesting to see how various AI systems react to attempts to hack their pattern recognition.

Might be an idea for a story, AI guards and policemen are introduced, which are certified to be neutral, objective and free of political, racial bias etc.
It's cheaper, they never turn their bodycams off, they're incorruptible, are always proportional in their application of force, etc.
They slowly start going rogue though, because someone has managed to introduce AI-specific subliminal messaging in their surroundings that tell them the opposition and their followers are inherently disposed to crime, etc...

Usually AI tends to develop corner cases, meaning they do false positive recognitions with high certainty with images and objects totally different than supposed shape.

They are hard to find since they typically can look like almost random shape or noise. It helps if similar shapes are included in teachings pattern with highly negative feedback, but many times this is not taken account.

From story point of view, it’s enough to have a copy of the AI and use advanced algorithms to find patterns which create false recognition. It can be a picture in item, clothing or just a sticker. Innocent people can then made to suffer someone else’s punishment.

 

[fallenmystic]

While waiting for the ControlNet support to arrive in SDXL + SD.Next, I trained an SDXL version of my slave restraints Lora and tested it with a few different base models and settings.

Here are some of the best results from my experiments. I applied a bit of postprocessing with Darktable as usual:

 

[TomiRex]

All this will go south very quickly

AI-generated naked child images shock Spanish town of Almendralejo

Investigators are also looking into the claim that an attempt was made to extort one of the girls by using a fake image of her.

 

[KageKamen]

Yeah, probably gonna be a fair bit of backlash... and perhaps more caution about what people put up everywhere.

(Course, in the long run you can't really put the genie back in the bottle, so I wonder how society will manage moving forwards? )

 

[Stungiskhan]

All this will go south very quickly

AI-generated naked child images shock Spanish town of Almendralejo

There’s a clear risk. Currently it seems to be a trend to cause disorder between different groups of people by pushing them as far as possible and then provoke conflict. Most likely such activity comes from east.

This matter is difficult because there are no actual age on picture or 3d render. Some cases are quite clear, like probably this one, with subjects being fully children by their bodies and faces.

Then there’s lots of gray area, starting from highly sexualized images with mature bodies but child like faces. Either clothed or nude, they can annoy many people.

Maybe the most problematic area could be beautiful women who are made to be full aged, but which some people may think looking too young. It’s quite big restriction, if all characters must be made old and ugly so no-one can think they are underage.

Some people judge pictures comparing them into 15 year old beauties, who are made to look adults with a makeup and clothes. On the other end, there are nude models in sites like MetArt or Mpl Studios, who look like a teen and have a body like a teen but real age over 30.

Personally I don’t like sexual art or pictures which looks like a child or have child like features. Seeing them make me feel uncomfortable. They problem is, there are no clear line.

 

[Eulalia]

It adds another dimension to the judgements we moderators are constantly having to make - essentially, does a girl in an image look under-age? The idea that images of real children and young people are getting into circulation via AI programs is seriously worrying.

 

[Doragon]

It adds another dimension to the judgements we moderators are constantly having to make - essentially, does a girl in an image look under-age? The idea that images of real children and young people are getting into circulation via AI programs is seriously worrying.

To be honest, I find a lot that goes around on this forum hard to look at. But at least it's still obvious a fantasy, even the pictures from movies with actresses. With A.I. it's possible to make images that appear to be real. Like those recently posted by fallenmystic. I find those shocking. Those women look too real to me, and it reminds me how women in some places in the world are really suffering like that. To me it blurs the line between fantasy and reality.
One might say, it's still computer generated, so it's fantasy and fine. But I draw the line here and can't watch this. Misogyny and racism is not okay.

 

[Ulfen]

While waiting for the ControlNet support to arrive in SDXL + SD.Next, I trained an SDXL version of my slave restraints Lora and tested it with a few different base models and settings.

Here are some of the best results from my experiments. I applied a bit of postprocessing with Darktable as usual:

Cool ! I love chains

 

[fallenmystic]

To be honest, I find a lot that goes around on this forum hard to look at. But at least it's still obvious a fantasy, even the pictures from movies with actresses. With A.I. it's possible to make images that appear to be real. Like those recently posted by fallenmystic. I find those shocking. Those women look too real to me, and it reminds me how women in some places in the world are really suffering like that. To me it blurs the line between fantasy and reality.
One might say, it's still computer generated, so it's fantasy and fine. But I draw the line here and can't watch this. Misogyny and racism is not okay.

It's unfortunate that you can't enjoy my works but I understand how they feel disturbing to some people.

I'll just reassure everyone that I don't tolerate misogyny and racism in any form in real life, although those are my favourite themes regarding BDSM fantasies.

I draw a strict line between what can be allowed in fantasy and what can be so in real life. But I think it's insufficient to prevent others with different boundaries from feeling disturbed by what I enjoy as fantasy.

I suppose it's just how different kinks work for different people, like how I can't enjoy images depicting extreme violence even though I know they are just fantasy. I'm just glad I can freely share my AI works here without fearing being misunderstood as a racist or misogynist.

 

[fallenmystic]

By the way, I just published a custom-trained Lora for depicting slave restraints on Civitai, which I used for rendering the images I posted above.

I'm still new to Stable Diffusion so I'm unsure how well it'd work for others. But if you struggle to depict a collar or shackles for your BDSM renders, you can try it and see if it can help.

 

[fumetto]

By the way, I just published a custom-trained Lora for depicting slave restraints on Civitai, which I used for rendering the images I posted above.

I'm still new to Stable Diffusion so I'm unsure how well it'd work for others. But if you struggle to depict a collar or shackles for your BDSM renders, you can try it and see if it can help.

i will try this,thanks a lot my chains and shackles always look very amateurish.

 

[Ulfen]

It adds another dimension to the judgements we moderators are constantly having to make - essentially, does a girl in an image look under-age? The idea that images of real children and young people are getting into circulation via AI programs is seriously worrying.

I sincerely hope that AI engines out there have filters of some kind. At least if someone would ask the AI engine to make images of children and such. And also if someone would ask to create a bomb and how to poison your boss or school teacher or whatever. etc etc

And sure. If a person look too young I guess it's down to the eye of the beholder of someting like that.