91–100 of 519 posts
Earlier quoted context omitted.
The future is probably models trained almost exclusively on porn.
Porn has driven many tech advances. I predict that models trained on specific porn genres will appear as soon as training a good model is doable for under $5000. They’ll get here much quicker if we get video to that mark first.
Earlier quoted context omitted.
I don't understand why so many people call Stable Diffusion open source.
You can download the model weights and run them offline. At least, you could in v1.4. I assume this is still possible on v2.0?
Is there a good explanation of how to train this from scratch with a custom dataset[0]? I've been looking around the documentation on Huggingface, but all I could find was either how to train unconditional U-Nets[1], or how to use the pretrained Stable Diffusion model to process image prompts (which I already know how to do). Writing a training loop for CLIP manually wound up with me banging against all sorts of stra…
There is working training code for openCLIP https://github.com/mlfoundations/open_clip
But training multi-modal text-to-image models is still a _very_ new thing, in terms of the software world. Given that, my experience has been that it's never been easier to get to work on this stuff from the software POV. The hardware is the tricky bit (and preventing bandwidth issues on distributed systems).
That isn't to say that there isn't code out there for training. Just that you're going to run into issues and learning how to solve those issues as you encounter them is going to be a highly valuable skill soon.
edit:
I'm seeing in a sibling comment that you're hoping to train your own model from scratch on a single GPU. Currently, at least, scaling laws for transformers [0] mean that the only models that perform much of anything at all need a lot of parameters. The bigger the better - as far as we can tell.
Very simply - researchers start by making a model big enough to fill a single GPU. Then, they replicate the model across hundreds/thousands of GPU's, but feed each on a different set of the data. Model updates are then synchronized, hopefully taking advantage of some sort of pipelining to avoid bottlenecks. This is referred to as data-parallel.
They apparently tried to combat NSFW generation by filtering the training dataset not to include any.
I guess SD is betting on saving $ on compute being more important in this space than the ability to gatekeep certain queries. And the tradeoff is that you need to do nsfw filtering in your released model.
It will be interesting to see who's right in 2 years.
Highlights: 768x768 native models (v1.x maxed out at 512x512) a built-in 4x upscaler: "Combined with our text-to-image models, Stable Diffusion 2.0 can now generate images with resolutions of 2048x2048–or even higher." Depth-to-Image Diffusion Model: "infers the depth of an input image, and then generates new images using both the text and depth information." Depth-to-Image can offer all sorts of new creative applica…
Not really surprised they did this, but be sure some communities will have it fine tuned on porn now-ish. So probably they did it for legal reasons in case illegal materials are generated and they are real companies/people with their names on the release?
Earlier quoted context omitted.
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?
Without the changes they made to Stable Diffusion, it was already able to generate CP. That's why they restricted it from doing so. It did not have child pornography in the training set, but it did have plenty of normal adult nudity, adult pornography, and plenty of fully clothed children, and was able to extrapolate. Anyway, one obvious application: FBI could run a darknet honeypot site selling AI-generated child po…
Earlier quoted context omitted.
I am not clicking that link because no one should take the risk of you proving your point of what horrors could pop out of one of these models. I will say that while the government backlash is inevitable just like it was with encryption, these image generation models are so easy to train on consumer hardware that the cat is hopelessly out of the bag. It might as well be thoughtcrime.
Link doesn't show any model output - it's an screenshot of photoshop refusing to edit a banknote.
Earlier quoted context omitted.
Quoted post unavailable.
This comment pretty clearly breaks the commenting guidelines https://news.ycombinator.com/newsguidelines.html >Be kind. Don't be snarky. Have curious conversation; don't cross-examine. Please don't fulminate. Please don't sneer, including at the rest of the community. Edit out swipes. Comments should get more thoughtful and substantive, not less, as a topic gets more divisive.
Earlier quoted context omitted.
In practice, it's unclear how well avoiding training on NSFW images will work: the original LAION-400M dataset used for both SD versions did filter out some of the NSFW stuff, and it appears SD 2.0 filters out a bit more. The use of OpenCLIP in SD 2.0 may also prevent some leakage of NSFW textual concepts compared to OpenAI's CLIP. It will, however, definitely not affect the more-common use case of anime women with v…
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…
The model does not contain the images themselves though. I think it would not be classified as that.