Stable Diffusion 2.0 • Hacker News Board

Re: Stable Diffusion 2.0

#161

Post by millimeterman » Thu, Nov 24, 2022, 5:39 AM UTC

liuliu wrote: post #41

Seems the structure of UNet hasn't changed other than the text encoder input (768 to 1024). The biggest change is on the text encoder, switched from ViT-L14 to ViT-H14 and fine-tuned based on https://arxiv.org/pdf/2109.01903.pdf . Seems the 768-v model, if used properly, can substantially speed-up the generation, but not exactly sure yet. Seems straightforward to switch to 512-base model for my app next week.

I'm disappointed they didn't push parameter count higher, but I suppose they want to maintain the ability to run on older/lower end consumer GPUs. Unfortunately it severely limits how high-quality the output can be.

Re: Stable Diffusion 2.0

#162

Post by abeppu » Thu, Nov 24, 2022, 5:42 AM UTC

wodenokoto wrote: post #146

tamersalama wrote: post #107

Earlier quoted context omitted.

I agree that this a big wave, but I'm still struggling to find commercial (read: large organizations) applications.

Built a plugin for Power Point and sell it corporate wide.

Does it actually make anything in the corporate world better to use generated images in slides? When coworkers use stock photos which were presumably made by humans operating actual cameras, I don't think it's clear that their presentation is actually more valuable as a result.

Re: Stable Diffusion 2.0

#163

Post by ansgri » Thu, Nov 24, 2022, 5:42 AM UTC

prawn wrote: post #131

Hopefully related: If I'm a photographer wanting to improve resolution of my content for printing, what's my current best bet for upscaling? Is it realistic to make use of this on the command line, feeding it my own images? Or has someone wrapped it in an app or online service?

As a counter-recommendation, Topaz’s much-advertised Gigapixel AI is rarely useful. Their Denoise and Sharpen apps are good though.

Re: Stable Diffusion 2.0

#164

Post by simion314 » Thu, Nov 24, 2022, 5:44 AM UTC

bloaf wrote: post #71

jimbob45 wrote: post #67

Earlier quoted context omitted.

Banknote printing is primarily protected against on the hardware level of printers, no? With the nigh-invisible unique watermark left by every printer, there’s virtually no way you’d get away with it. My guess is that the Photoshop filter exists mostly as a barrier against the crime of convenience.

My point is that there is precedent for governments requiring companies to implement restrictions on what images can be handled by their software. As I explained: This kind of mandated restriction is looming over AI. Companies are trying to get out in front of these restrictions so they can implement them on their own terms.

>My point is that there is precedent for governments requiring companies to implement restrictions on what images can be handled by their software

But images of boobs are still legal. So this NSFW filter seems to be much more above then the law asks. Is the issue is that even if you do not train with CP you might get the model so output something that some random person will get offended and label it as CP? I assume that other companies can focus on NSFW and have their lawyers figure this out, IMo would be cool that someone sues the governments and make them reveal facts about their concern that CP of fake or cartoon people is dangerous" , I think they could focus on saving real children then cartoon ones.

Re: Stable Diffusion 2.0

#165

Post by ericd » Thu, Nov 24, 2022, 5:44 AM UTC

LASR wrote: post #120

searchableguy wrote: post #30

Earlier quoted context omitted.

What are you building?

It started as an AI-powered MS paint for my son. But after demoing it to a few coworkers, it morphed into a bit more than that. Now it’s more of a storybook creator that young kids can use to generate their own stories. Not looking to monetize at all. But inference is expensive. So might have something to cover costs. Some backstory: When I was growing up in the early 90s, my dad took me into his office over the week…

I have some young kids, would love to try this out with them if you’re sharing.

Re: Stable Diffusion 2.0

#166

Post by pkdpic » Thu, Nov 24, 2022, 5:48 AM UTC

kristopolous wrote: post #25

rogers18445 wrote: post #3

They apparently tried to combat NSFW generation by filtering the training dataset not to include any.

Did they exclude celebrities, politicians, and religious and political symbols? Deceitful extremists and vengeful criminals fabricating lies seem to be a far more serious problem than fantasy porno.

That's a really interesting point, and it makes me realize that the Nancy Reagan 'what constitutes porn' question is obviously super old and problematic.

Also lexica.art is swarming with celebrity fantasy porn that just has a thin stylistic filter of paintings from the 19th century. And a plethora of furry daddies that you can't not love.

I get why these models should be curated but I also like that the sketchy porn possibilities keep them feeling un-padded / interesting / dangerous.

Then again this all is probably really dangerous so maybe that's silly.

Re: Stable Diffusion 2.0

#167

Post by speedgoose » Thu, Nov 24, 2022, 5:49 AM UTC

vsskanth wrote: post #77

Can this transform an image into a vector illustration ?

img2img can change the style of an image so you can make it look like a vector illustration but it will still be a bitmap.

Re: Stable Diffusion 2.0

#168

Post by radu_floricica » Thu, Nov 24, 2022, 5:52 AM UTC

Speaking of business models for AI, and the fact that stable diffusion is anti-trained for porn. Somebody with an old terra byte image porn collection right now: "Hold my beer, my time has come!"

Re: Stable Diffusion 2.0

#169

Post by the_shivers » Thu, Nov 24, 2022, 5:58 AM UTC

joshspankit wrote: post #68

Gigachad wrote: post #11

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.

What tech advances would those be?

Re: Stable Diffusion 2.0

#170

Post by corysama » Thu, Nov 24, 2022, 5:59 AM UTC

sorenjan wrote: post #133

I've seen references to merging models together to be able to generate new kinds of imagery or styles, how does that work? I think you use Dreambooth to make specialized models, and I think I got an idea about how that basically assigns a name to a vector in the latent space that represents the thing you want to generate new imagery of, but can you generate multiple models and blend them together? Edit: Looks like AU…

It’s my understanding that, amazingly enough, blending the models is done by literally performing a trivial linear blend of the raw numbers in the model files.

Someone even figured out they could get great compression of specialized model files by first subtracting the base model from the specialized model (using plain arithmetic) before zipping it. Of course, you need the same base file handy when you go to reverse the process.