Hacker News Reader

Comment by viraptor | Hacker News Reader

Just wait some time. People release SD loras all the time. Once SD3 is open, you'll be able to get a patched model in days/weeks.

View on HN

A blogger I follow had an article explaining that the NSFW models for SDXL, are just now SORT OF coming up to the quality of SD1.5 “pre safety” models.

It’s been 6 months and it still isn’t there. SD3 is going to be quite awhile if they’re baking “safety” in even harder.

1.5 is still more popular than xl and 2 for reasons unrelated to safety. The size and generation speed matter a lot. This is just a matter of practical usability, not some idea of the model being locked down. Feed it enough porn and you'll get porn out of it. If people have incentive to do that (better results than 1.5), it really will happen within days.

Due to the pony community the SDXL nsfw models are far superior to SD1.5. Only issue is that controlnets don’t work with that pony SDXL fine tune

I am slightly aware of the pony models.

I wish I had something more clever to comment on it. I know what they’re doing which is cool and why which is, IDK, live and let live and enjoy your own kink. It just a little funny some of the most work put into in the fine tuning models.. is from the pony community.

So all I have is…

It's not just them. For example, 4chan is a surprisingly good way to get the most recent scoop on good models (both text and images), setup guides etc - if you can tolerate the inevitable, well, 4chan-ness of it. And the reason is exactly the same: a lot of people there really, really, really, want to generate porn and to chat with sexbots, and they're putting a lot of effort into getting the best (and least censored) results out of the resources that they have.

4Chan was/is an amazing resource for finding ways to bypass Dalle restrictions.

Sadly the elite ai hackers have shuffled off into private discords.

> IDK

We are just trying to satisfy your values though ponies and friendship.

Apparently most stuff that relies on the weights being resonably similar to SDXL don't work - control nets, LORAs, commonly-used inpainting patches, the lot. It seems to go well beyond fine tuning and be substantially retrained to the point it's a good chunk of the way to being a different model entirely, and the amount of training time is on the order of what went into SDXL originally too from what I can tell.

I hope they call it Horse Diffusion.

Reply on news.ycombinator.com