Stable Diffusion (like Dall-E but open source and better)

Stable Diffusion now has a public release. One of the problems (or looked at another way, security features) of past AI image generator models is that they are tightly controlled by the companies that created them. You either get very low-res versions of the models for free (e.g. DALL-E mini) or tight restrictions over the number of images (e.g. midjourney).

Stable diffusion is open source and there are no restrictions/limitations on its use. Of course, the downside is you need to run the model locally and flex that GPU. But since this is a gaming forum, I doubt that will be an issue for many people here!


Website post:

Installation guide that I used (disclaimer: I don’t know the author and use at your own risk, but all looks fine to me):

Once I get it working I will start posting some examples here…

Here is a link that gives you some ideas of prompts and how they affect the image with stable diffusion:

So, the real question - does this new method allow for the nsfw outputs, or not?

My friend, who has tried such things, says yes it does.

My friend says excellent

So, yeah, it’s 2024, the timeline sucks, truth is a concept that died somewhere in the late 2010s, and little old me finally got around to dabble a bit in AI generation.

Well. It’s gotten pretty good. Doubly so if you consider that this is basically an algorithm applying learned patterns to white noise to filter out a desired image. It’s about 2 steps away from black magic voodoo tech.

The newer models (SDXL and, if you’re into non-realistic stuff, PonyXL) come with a somewhat understandable prompt syntax and less baggage you have to put into the negatives. PonyXL and its derivates, in particular, recognizes a lot of concepts (something the older models are sorely lacking in…).

Given that they were trained on some rather…illustrious websites, many of those concepts are decidedly NSFW.

And yes, I’m running a local copy - I want to see what it can do, no trying to get around various filters to keep stuff family friendly. And I admit…the results were impressive. I’m rather happy that most users are just pervy horndogs, and not political activists. I assume with a few dozen fotos of a public figure and some imagination you can produce some rather questionable images.

Anyways, enough doom and gloom. Here’s a cute goblin doing goblin things. Sleeping and eating. And feeding the “amazing chest ahead”.

(No goblins were harmed during the making of these pictures. However, I’m still a few hamburgers short for some reason.)

Apparently Stable Diffusion 3 has fixed the text problem, at least based on the examples in the research paper. We’ll have to see how it really performs without cherrypicking.