Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Stable Diffusion 2.0

stability.ai

Post reply on HN
Print view View original

261–270 of 519 posts

Re: Stable Diffusion 2.0

#261

Post by theclansman »

fastball wrote: post #76
kmeisthax wrote: post #45

Earlier quoted context omitted.

The main reason why Stable Diffusion is worried about NSFW is that people will use it to generate disgusting amounts of CSAM. If LAION-5B or OpenAI's CLIP have ever seen CSAM - and given how these datasets are literally just scraped off the Internet, they have - then they're technically distributing it. Imagine the "AI is just copying bits of other people's art" argument, except instead of statutory damages of up to…

Is artificially generated CSAM that doesn't actually involve children in its production not an improvement over the status quo?

I remember Louis CK made a joke about this, in regards to pedophiles (who are also rapists), what are we doing to prevent this? Is anyone making very realistic sex dolls that look like children? "Ew no that's creepy" well I guess you would rather them fuck your children instead. It's one of those issues that you have to be careful not get too close to, because you get accused by proximity, if you suggest something like what I said before people might think you're a pedophile. So in that way, nobody wants to do anything about it.

Re: Stable Diffusion 2.0

#262

Post by concordDance »

braingenious wrote: post #72
datalopers wrote: post #40

Earlier quoted context omitted.

They’ve ensured the only way to create CSAM is through old-fashioned child exploitation, meanwhile all perfectly humane art and photography is at risk of AI replacement. This is a huge missed opportunity to actually help society.

LMFAO What do you propose? The FBI releases a CSAM data set for devs to use for “training”? Would you be the one to create the model? Would you run a business that sells synthetic CSAM?

Yes to the first and no to the second seem the obvious answers here.

Re: Stable Diffusion 2.0

#264

Post by jeroenhd »

Gigachad wrote: post #171
tempthrowaway11 wrote: post #83

Earlier quoted context omitted.

They’re already out there, although they’re hard to find via Google - people are doing wild things like “merging” hentai models with models trained on real life porn to get realistic poses and lighting with impossible anatomy. The scary thing is that you can then train it further with things like DreamBooth to start producing porn of celebrities… or, even more worrying, people you know. Seriously folks, we are within…

I have no idea how it works but I have seen people talking about models trained to draw furry art. And I assume no one spent the millions on AWS to train a full model from scratch.

I believe what they do is take the released version of stable diffusion and then continue training from there with their own image sets. I came across their attempts when looking into how to train the model based on some images of my own; their data set so far reaches between tens of thousands and hundreds of thousands images.

All the difficult parts (poses, backgrounds, art styles) has already been done by the SD researchers, the porn network only needs reference material for the NSFW description/tags/details. This is significantly cheaper.

A similar project, training SD to output images in the style of Arcane, is incredibly successful in replicating the animation style with what seems to be very little actual training data.

I don't think you need to start from scratch at all if you use the SD model as a base, all you need to do is to train it on specific concepts, styles and key words that the original doesn't have.

Re: Stable Diffusion 2.0

#265

Post by wokwokwok »

CamperBob2 wrote: post #211
wokwokwok wrote: post #175

Earlier quoted context omitted.

Quite right, but… > That argument also makes little sense when you consider that the model is a couple gigabytes itself, it can't memorize 240TB of data, so it "learned". The matter is really very nuanced and trivialising it that way is unhelpful. If I recompress 240TB as super low quality jpgs and manage to zip them up as single file that is significantly smaller than 240TB (because you can), does the fact they are…

If I recompress 240TB as super low quality jpgs and manage to zip them up as single file that is significantly smaller than 240TB (because you can), does the fact they are not pixel perfect matches for the original images mean you’re not violating copyright? If you compress them down to two or three bytes each, which is what the process effectively does, then yes, I would argue that we stand to lose a LOT as a techno…

Maybe?

Does that mean it’s worthless to try to train an ethical art model?

Is it not helpful to show that you can train a model that can generate art without training it on copyrighted material?

Maybe it’s good. Maybe not. Who cares if people waste their money doing it? Why do you care?

It certainly feels awfully convenient for that there are no ethically trained models because it means no one can say “you should be using these; you have a choice to do the right thing, if you want to”.

I’m not judging; but what I will say is that there’s only one benefit in trying to avoid and discourage people training ethical models:

…and that is the benefit of people currently making and using unethically trained models.

Re: Stable Diffusion 2.0

#266

Post by jeroenhd »

NKosmatos wrote: post #199

Wow, just wow! Newbie question, why can’t someone just take a pre-trained model/network with all the settings/weights/whatever and run it on a different configuration (at a heavily reduced speed)? Isn’t it like a Blender/3D studio/Autocad file, where you can take the original 3D model and then render it using your own hardware? With my single GOU it will take days to raytrace a big scene, whereas someone with multipl…

The main limitation for running these AIs is that you need tons of VRAM available for your GPU to get any good performance out of them. I don't have a video card with 12GiB of VRAM and I don't know anyone who does.

If you're willing to wait more (30 seconds per image, assuming limited image sizes) there are repositories that will run the model on the CPU instead, leveraging your much cheaper RAM.

In theory you could swap VRAM in and out in the middle of the rendering process, but this would make the entire process incredibly slow. I think you'll have more success just running the CPU version if you're willing to accept slowdowns.

Re: Stable Diffusion 2.0

#267

Post by ArnoVW »

coldblues wrote: post #137

I just thought about this, so bare in mind that I don't know much of the technical implications of this, but: Couldn't we train a very good model by distributing the dataset along with the computing power using something similar to folding@home?

The limit for this sort of exercise is "holding everything in memory". Because training of neural networks require that one updates the weights frequently. An NVIDI A100 has a bandwidth of 2 Tb/sec. Your home ADSL something in the order of 10 Mbit. And then there's latency.

Mind you, theoretically that is a limitation of our current network architectures. If we could conceive a learning approach that was localised, to the point of being "embarrassingly parallel", perhaps. It would probably be less efficient, but if it is sufficiently parallel to compensate for Amdahl's law, who knows?

Less theoretically, one could imagine that we use the same approach that we use in systems engineering in general: functional decomposition. Instead of having one Huge Model To Rule Them All, train separate models that each perform a specific, modular function, and then integrate them.

In a sense this is what is currently happening already. Stable Diffusion have one model to generate img2depth, to generate an estimation which parts of a picture are far away from the lense. They have another model to upscale low res images to high res images, etc etc. This is also how the brain works.

But it is difficult to see how this sort of approach could be applied to very small scale, low contextual tasks, like folding@home.

Re: Stable Diffusion 2.0

#269

Post by Scandiravian »

What's the potential of using this for image restoration? I've been looking into this recently as I've found a ton of old family photos, that I'd like to digitize and repair some of the damage on them

There are a lot of tools available, but I haven't found anything where the result isn't just another kind of bad, so if the upscaling and inference in this model is good, it should in theory be possible to restore images by using the old photos as the seed, right?

Re: Stable Diffusion 2.0

#270

Post by jeroenhd »

charcircuit wrote: post #87

I dislike how they call their model open source even though there are restrictions on how you can use the model. The ability to use code however you want and not have to worry about if all the code you are using is compatible with your use case is a key part of open source.

The code is open source, the model is a data file that the open source code operates on. It's similar to engine recreations for old games (OpenRCT, OpenTTD) that use original, proprietary assets to play the games with their open source engines.

Similar to those games, anyone is also able to distribute their own open data files if they so wish It's unlikely anyone actually will start training an open source AI model from scratch because doing so costs insane amounts of money, but the same can be said about the many hours of work recreating game assets can take for open source game engines.

Post reply on HN