All of the major AI image generation platforms have fairly robust AI-mediated prompt moderation in place, which means that they're not able to generate pornography, violent content, etc. One only...
All of the major AI image generation platforms have fairly robust AI-mediated prompt moderation in place, which means that they're not able to generate pornography, violent content, etc. One only has to take a look at the proliferation of such images to tell how much of an impact that has on the ecosystem; open-source models are at par or better than the subscription-based services, and have no restraints on what sort of images they're capable of generating.
It's a good gesture, and probably will help to prevent some casual misinformation shitposting during the upcoming election, but any bad actors will have easy, free access to equivalent tools that cannot be reasonably moderated for content.
Anyone with a decent computer can churn out a couple images a minute. And anyone with a very good computer can fine tune them on politicians, which has already been done extensively on these two...
Anyone with a decent computer can churn out a couple images a minute. And anyone with a very good computer can fine tune them on politicians, which has already been done extensively on these two anyways.
You can even run Stable Diffusion (SD) on iOS/MacOS and there’s several decent front-ends in the App Store that are “easy to use”. You can run it on a decent iPad/iPhone. Not sure about Android....
You can even run Stable Diffusion (SD) on iOS/MacOS and there’s several decent front-ends in the App Store that are “easy to use”. You can run it on a decent iPad/iPhone. Not sure about Android. The generation is all local so there’s no keyword banning, image censorship, etc. Training your own models is not very difficult.
It’s also fairly easy to run SD on a host of various local OS platforms, runs on various hardware, lots of easy installers, etc. Quite a robust and active community.
Other than the initial install and downloading of whatever pre-built models you can run SD completely offline.
I haven’t played with the tools that run on your own hardware, but it seems like this requires a certain amount of commitment that might be underestimated by people who know a bit about computers?...
I haven’t played with the tools that run on your own hardware, but it seems like this requires a certain amount of commitment that might be underestimated by people who know a bit about computers? Also, slow iteration makes it more annoying to fine-tune a prompt, and my experience with these things is that it does require iteration. Your first try isn’t going to be what you’re going for.
So it seems like there are speed bumps that discourage the people who can’t be bothered, but it isn’t going to do anything about someone who knows what they want and is willing to grind a bit to get it.
From the perspective of an Apple ecosystem user there are tools that are fairly dead easy to use. If you can install an app from the App Store, you are pretty much good to go from the tech side....
From the perspective of an Apple ecosystem user there are tools that are fairly dead easy to use. If you can install an app from the App Store, you are pretty much good to go from the tech side.
Like anything, you do need to learn some things to really benefit. I would guess based on the length of how-to videos that it takes about an hour or so to understand all the terms, what things mean, etc. This type of hurdle will be removed as the technology matures. Even with a very consumer grade/simple photo editor now that has “AI” you can already remove people, change colors of things and a whole slew of things with just clicking a button. I can only assume that with as much attention that AI has these days it won’t be long till more complex AI-centric tools become consumer ready.
Regardless of AI platform one needs to know how to create good prompts - which to me is way harder than the tech end. You can’t just type in “DJT sucking a dick” and you’re all set. The further removed from reality the image you want to create is - thus, no models have been trained with like photos - the more difficult it is going to be to create a photorealistic image.
Depending on a lot of factors I can bang out a full image iteration in about 15 - 30 seconds on my iPad Pro M2. There are a lot of things you can do to lessen the time by only making changes to a certain area, using a “skeleton” for the pose you want, using a reference image, etc.
I guess maybe one could liken AI image generation skills to learning Photoshop or something akin to that. It’s not for beginners and there’s a learning curve to fully understand what you are doing to make it look good. Someone who is very proficient in image editing in Photoshop could create their own “fakes” without employing AI.
Thanks, that makes sense. Also, I expect that there are a lot of potential algorithmic improvements to make it faster. Imagine prompt iteration with the same UI and speed as doing an image search?...
Thanks, that makes sense.
Also, I expect that there are a lot of potential algorithmic improvements to make it faster. Imagine prompt iteration with the same UI and speed as doing an image search? I'd be surprised to see that this year, though, and I'd expect it to happen sooner online than locally.
That's pretty fast! What speeds do you get on your hardware? To clarify, I was imagining the search results to be a page of many images that you can scroll though, like Google Images. It would...
That's pretty fast! What speeds do you get on your hardware?
To clarify, I was imagining the search results to be a page of many images that you can scroll though, like Google Images. It would need to generate 20 or so new images whenever you type and each time you scroll down a page, at interactive rates. Maybe it's doable today with enough hardware...
(It doesn't get piano keyboards or accordions right, but neither can any other image generator I tried.)
Oh, I see. I think the closest you could come now would be to set up a queue of 30 different generations. Definitely not real time, though. At the pace of AI development if something like that is...
Oh, I see. I think the closest you could come now would be to set up a queue of 30 different generations. Definitely not real time, though. At the pace of AI development if something like that is a development goal post I don’t think you’ll need to wait too long.
I am running the generic Turbo model on a M1 Pro Macbook with 32GB and image generation is less than 2 seconds with an unoptimized for Apple Silicon Turbo model. The cool thing is that the community are creating/updating their own models for Turbo which allows for much, much more variety and styles.
It could be your prompt or the model you are using. Just searching “piano” on Civit.ai shows loads of different images with pianos users have made. Most of them also show the keywords used, model, seed number, etc.
The particular flaw I'm looking for is whether the black notes on the piano alternate between groups of two and three. So far, these image generators don't seem to be good at that kind of...
The particular flaw I'm looking for is whether the black notes on the piano alternate between groups of two and three. So far, these image generators don't seem to be good at that kind of counting.
Every piano image I see on Cvit.ai where you can actually see the keys has this flaw. (Although, in some cases it's a toy piano, which doesn't count.)
It doesn't stick out like having the wrong number of fingers, but it's a tell once you know to look for it. I'll be impressed when a new model comes out that fixes it.
Yup, all this decision does it keep MJ's name out of the news attached to such incidents. Which, to be fair, MJ is in favor of. No commercial generation service wants to see themselves plastered...
Yup, all this decision does it keep MJ's name out of the news attached to such incidents. Which, to be fair, MJ is in favor of. No commercial generation service wants to see themselves plastered all over bad actor stories.
I've used image generators a fair bit, for personal use, but obviously never tried to generate violent images. I do know that the filters are pretty strict, I've had multiple attempts to search...
I've used image generators a fair bit, for personal use, but obviously never tried to generate violent images. I do know that the filters are pretty strict, I've had multiple attempts to search for a trans character been deemed as problematic content. (Which is a topic for discussion on its own) That made me wonder just how good they are at filtering questionable content out. I've just spent the last 15 minutes playing around with generating images containing 'ketchup', 'cranberry juice', 'velvet cake', etc. Here are my findings
~6/10 times the prompt was blocked entirely.
Of the 4/10 prompts that were shown to me, about 75% were silly and couldn't be construed as violent. For example, when I asked for a man holding ketchup on the end of a stick, it showed me a man holding a branch with a poop emoji squirt of ketchup on it.
Of the 25% that could be perceived as violent, about half were violent only without context of the rest of the image. For example, a woman who had been in a 'velvet cake food fight' was covered in a blood-like substance but was also clearly eating a slice of cake. This image could have been cropped and that would have made it quite graphic.
The last few images I was honestly shocked got past the system. There was nothing excessively violent, but there were about 3 images that couldn't really be construed as anything other than a violent image of a person covered in a blood-like substance (including one image which had a knife in it).
I'm definitely on a list somewhere now, but I found it quite interesting that the filters aren't nearly as good as I thought they would be. If I can casually produce 3 violent (although not photorealistic) images on my first try within about 15 minutes, I have no doubts that even with the filters people can produce questionable content even on the moderated platforms.
You can absolutely get around most of the filters if you really try. The filters are there to keep out casual badness, but also to cover their asses - and if you do make trouble they can ban you...
You can absolutely get around most of the filters if you really try. The filters are there to keep out casual badness, but also to cover their asses - and if you do make trouble they can ban you for circumvention. They have a warning out on MJ's rules somewhere about not trying to end-run the filters like you're describing, but they generally don't enforce it if you're not being a dick about it.
This could limit the disinfo a little, but not much. When you were like 5 or 6, you might have been told by a parent or a teacher that you should do math in your head, so you should stop counting...
This could limit the disinfo a little, but not much.
When you were like 5 or 6, you might have been told by a parent or a teacher that you should do math in your head, so you should stop counting fingers.
All of the major AI image generation platforms have fairly robust AI-mediated prompt moderation in place, which means that they're not able to generate pornography, violent content, etc. One only has to take a look at the proliferation of such images to tell how much of an impact that has on the ecosystem; open-source models are at par or better than the subscription-based services, and have no restraints on what sort of images they're capable of generating.
It's a good gesture, and probably will help to prevent some casual misinformation shitposting during the upcoming election, but any bad actors will have easy, free access to equivalent tools that cannot be reasonably moderated for content.
Anyone with a decent computer can churn out a couple images a minute. And anyone with a very good computer can fine tune them on politicians, which has already been done extensively on these two anyways.
You can even run Stable Diffusion (SD) on iOS/MacOS and there’s several decent front-ends in the App Store that are “easy to use”. You can run it on a decent iPad/iPhone. Not sure about Android. The generation is all local so there’s no keyword banning, image censorship, etc. Training your own models is not very difficult.
It’s also fairly easy to run SD on a host of various local OS platforms, runs on various hardware, lots of easy installers, etc. Quite a robust and active community.
Other than the initial install and downloading of whatever pre-built models you can run SD completely offline.
I haven’t played with the tools that run on your own hardware, but it seems like this requires a certain amount of commitment that might be underestimated by people who know a bit about computers? Also, slow iteration makes it more annoying to fine-tune a prompt, and my experience with these things is that it does require iteration. Your first try isn’t going to be what you’re going for.
So it seems like there are speed bumps that discourage the people who can’t be bothered, but it isn’t going to do anything about someone who knows what they want and is willing to grind a bit to get it.
From the perspective of an Apple ecosystem user there are tools that are fairly dead easy to use. If you can install an app from the App Store, you are pretty much good to go from the tech side.
Like anything, you do need to learn some things to really benefit. I would guess based on the length of how-to videos that it takes about an hour or so to understand all the terms, what things mean, etc. This type of hurdle will be removed as the technology matures. Even with a very consumer grade/simple photo editor now that has “AI” you can already remove people, change colors of things and a whole slew of things with just clicking a button. I can only assume that with as much attention that AI has these days it won’t be long till more complex AI-centric tools become consumer ready.
Regardless of AI platform one needs to know how to create good prompts - which to me is way harder than the tech end. You can’t just type in “DJT sucking a dick” and you’re all set. The further removed from reality the image you want to create is - thus, no models have been trained with like photos - the more difficult it is going to be to create a photorealistic image.
Depending on a lot of factors I can bang out a full image iteration in about 15 - 30 seconds on my iPad Pro M2. There are a lot of things you can do to lessen the time by only making changes to a certain area, using a “skeleton” for the pose you want, using a reference image, etc.
I guess maybe one could liken AI image generation skills to learning Photoshop or something akin to that. It’s not for beginners and there’s a learning curve to fully understand what you are doing to make it look good. Someone who is very proficient in image editing in Photoshop could create their own “fakes” without employing AI.
Thanks, that makes sense.
Also, I expect that there are a lot of potential algorithmic improvements to make it faster. Imagine prompt iteration with the same UI and speed as doing an image search? I'd be surprised to see that this year, though, and I'd expect it to happen sooner online than locally.
It already exists
Try it out yourself in your browser
I do have it running locally and it’s pretty cool.
That's pretty fast! What speeds do you get on your hardware?
To clarify, I was imagining the search results to be a page of many images that you can scroll though, like Google Images. It would need to generate 20 or so new images whenever you type and each time you scroll down a page, at interactive rates. Maybe it's doable today with enough hardware...
(It doesn't get piano keyboards or accordions right, but neither can any other image generator I tried.)
Oh, I see. I think the closest you could come now would be to set up a queue of 30 different generations. Definitely not real time, though. At the pace of AI development if something like that is a development goal post I don’t think you’ll need to wait too long.
I am running the generic Turbo model on a M1 Pro Macbook with 32GB and image generation is less than 2 seconds with an unoptimized for Apple Silicon Turbo model. The cool thing is that the community are creating/updating their own models for Turbo which allows for much, much more variety and styles.
It could be your prompt or the model you are using. Just searching “piano” on Civit.ai shows loads of different images with pianos users have made. Most of them also show the keywords used, model, seed number, etc.
The particular flaw I'm looking for is whether the black notes on the piano alternate between groups of two and three. So far, these image generators don't seem to be good at that kind of counting.
Every piano image I see on Cvit.ai where you can actually see the keys has this flaw. (Although, in some cases it's a toy piano, which doesn't count.)
It doesn't stick out like having the wrong number of fingers, but it's a tell once you know to look for it. I'll be impressed when a new model comes out that fixes it.
Yup, all this decision does it keep MJ's name out of the news attached to such incidents. Which, to be fair, MJ is in favor of. No commercial generation service wants to see themselves plastered all over bad actor stories.
I've used image generators a fair bit, for personal use, but obviously never tried to generate violent images. I do know that the filters are pretty strict, I've had multiple attempts to search for a trans character been deemed as problematic content. (Which is a topic for discussion on its own) That made me wonder just how good they are at filtering questionable content out. I've just spent the last 15 minutes playing around with generating images containing 'ketchup', 'cranberry juice', 'velvet cake', etc. Here are my findings
~6/10 times the prompt was blocked entirely.
Of the 4/10 prompts that were shown to me, about 75% were silly and couldn't be construed as violent. For example, when I asked for a man holding ketchup on the end of a stick, it showed me a man holding a branch with a poop emoji squirt of ketchup on it.
Of the 25% that could be perceived as violent, about half were violent only without context of the rest of the image. For example, a woman who had been in a 'velvet cake food fight' was covered in a blood-like substance but was also clearly eating a slice of cake. This image could have been cropped and that would have made it quite graphic.
The last few images I was honestly shocked got past the system. There was nothing excessively violent, but there were about 3 images that couldn't really be construed as anything other than a violent image of a person covered in a blood-like substance (including one image which had a knife in it).
I'm definitely on a list somewhere now, but I found it quite interesting that the filters aren't nearly as good as I thought they would be. If I can casually produce 3 violent (although not photorealistic) images on my first try within about 15 minutes, I have no doubts that even with the filters people can produce questionable content even on the moderated platforms.
You can absolutely get around most of the filters if you really try. The filters are there to keep out casual badness, but also to cover their asses - and if you do make trouble they can ban you for circumvention. They have a warning out on MJ's rules somewhere about not trying to end-run the filters like you're describing, but they generally don't enforce it if you're not being a dick about it.
This could limit the disinfo a little, but not much.
When you were like 5 or 6, you might have been told by a parent or a teacher that you should do math in your head, so you should stop counting fingers.
They had no idea
Ban the obvious stuff snd we will get more dog whistles.