Show HN: Stable Diffusion Without Filters | Hacker News Reader

Show HN: Stable Diffusion Without Filters

healthydiffusion.com

52 pointssoheil34 comments

What’s your plan for this short-term and/or current long-term vision?

So, yet to have any prompts not load and confirmed NSFW prompts appear to work. Loading animation is way too much:

https://healthydiffusion.com/img/spinner-512.gif

No plans really, I was just very frustrated with not being able to quickly run a sentence through sd and get an image back even with the huge number of tools that already exist.

Understand, to be fair, not aware of any other service running with NSFW turned off, though have not really searched for it since you’re able to run it locally. Imagining running this for free is expensive though, you should just charge for it, or are least charge at cost; assuming there’s not other service with NSFW turned off.

Staying obvious, likely significant risk to running SD with NSFW turned off, but also not trying to run a service with it off.

I would also not be caching/hosting the outputs, just make it clear to the user it’s generated once and they need to save it locally, or it’s gone.

https://www.phantasmagoria.me has no filters, for what it's worth.

Well, technically it does, but you can see/use the image just fine, it just doesn't make it onto the front page.

Wants me to create an account first:(

Yeah, unfortunately GPUs are expensive :(

Agree, cheapest I have seen is $500 USD month, but also didn’t look that hard.

And yes, knew of phantasmagoria, but didn’t realize they had NSFW filters off; tested it, NSFW filters appear to be off.

I made this so I get it. I wonder if there is an alternate financial model that would make more sense than just straight up asking for money.

We should insert ads into the generated images. Imagine the Mona Lisa with a Coca Cola ad in the background.

Or have her wear Gucci.

No matter what I enter, the result is blank.

I tried with Chrome on Android Mobile.

Edit: All fixed, thank you both!

OP: How did you make this? Which version / flavor of SD is it using?

It seems to be slowing down or broken, it worked a few times and now the creation gif just spins forever. I anticipate an incoming HN hug of death.

"Right click" [1] image to view it in new tab. Had the same issue on MacOS Safari.

[1] tap and hold on Android, I would assume. Haven’t used one in years

Sorry an issue with ssl, try again.

What does “without filters” mean? The only filter I’m aware of is SD’s NSFW filter, which can be switched off anyway. Are there others?

No there aren't.

I assume it's more or less the same as as running sd on your pc, with the title being a contrast to other hosted services (which can be more restrictive about prompts).

Currently, https://healthydiffusion.com/recent contains a lot more nudity (and pepe the frog) than you usually see on these types of websites.

I think the site has been overloaded - that link just returns an empty page and new queries just spin forever.

Can’t use celebrity names either like Donald Trump

Donald Trump works, what prompts are you using? SD renders Trump oddly, same for all SD versions, sites, etc.

They’re just misinformed. I have seen lots of these “urban myths” from people trying to guess how the model works. In this instance, not realizing that there’s no explicit text-only filter; rather a vision-classifier based method that doesn’t explicitly encode any celebrities, but rather features relating to pornography, bigotry, etc.

There number of celebrities that render fine; as it relates to Trump, guessing it’s related to the images the model used.

Here are some examples, but obviously if you use the specific celebrity name you’re more likely to get more narrow results:

- https://lexica.art/?q=Celebrity

Here are Trump examples:

- https://lexica.art/?q=Trump

A bit off-topic, but I am mad impressed at how much lexica.art improved their performance in such a short amount of time (most likely, due to the recent funding they got). Their website used to be very laggy and slow on my pretty beefy desktop even for basic stuff like scrolling down, not even mentioning mobile performance.

Just opened the link on my phone again, and wow, it is blazing fast, even for non-trivial operations. Mad congrats to the team, as the product itself is extremely valuable.

Input “P = NP” for a proof. Computer scientists don’t want to you to know this one trick.

I get crappy results https://healthydiffusion.com/recent/5918210316301

The fabled poop algorithm, they said it was a pipe dream

Hugged to death? Seem to spin for quite a while. I'm using Edge on a PC.

Try again, hugged, but not dead yet.

Typed in "fat anime titties" and indeed got fat anime titties.

Can't complain about that.

What are you using on the backend?

Static results are served directly from Nginx including index.html (I was worried about hn hug of death) everything else is on unicorn/rails/sql/custom python scripts/etc.

So do you have the code for this? Where do you run it?

It worked but it took a few hours!

What do you mean, without filters? Stable diffusion does not have any filter as far as I know.

Oh, no I get it. It's true, the original stable diffusion repo does have filters, however there are many versions that you can run locally, usually with the filters turned off.

Reply on news.ycombinator.com