Series: AI Page 22 of 43

I swear I stumbled upon this innocently – at first – while making images of dystopian beach resorts for my next (unpublished) AI Lore book, Relaxatopia. And then, of course, like any good explorer I ruthlessly exploited what I found “for science.”

It seems I am not the only one who has accidentally discovered that Midjourney’s latest v6 Alpha release has very relaxed nudity filters relative to prior versions. And certainly I won’t be the last to get unasked for – but unintentionally appreciated – nudes from this model release.

I hesitate to even bring this to light publicly, because I don’t want them to change it. But they surely will, and they will surely do it whether it is me who posts about it or someone else. But I’ve realized something over the past couple years of experimenting with gen AI that these systems are quirky and glitchy af – especially when first released. So you have to get your kicks in early before the shit-house goes up in flames…

In any event, I put together an NSFW Medium post of 104 nudes and near nudes from Midjourney v6 Alpha release. Uploading large image sets to Medium is a PITA, so I hope they all got uploaded. There are actually more that I managed to get but didn’t include here. It’s very easy to create these once you understand the trick. That said, if you want gen AI nudes without all the hassle, do yourself a favor and download DiffusionBee for Mac and install the Stable Diffusion XL model. Quality is slightly lower, but it’s not so far off, and you can get away with a lot more.

There’s something fun about getting Midjourney to do these though when you know it’s not supposed to be doing it. It reminds me of being a teenager in the 90’s and watching scrambled Playboy channel on cable on the off-hope (sometimes rewarded) that you would catch a glimpse of something titillating.

I used to think it was somewhat weird that the first and primary use of a lot of human technology is pornography. But now I think that this is just Nature: it’s the Life Force welling up in all living things. It’s basic.

In Midjourney’s case, what it comes up with is actually quite beautiful (and some are disturbing in an intriguing way). Stable Diffusion XL seems to be much better for pure “smut,” but the Midjourney results I find in many cases to be quite artistic, and I feel like seeing them together is somehow liberating after all the wrist-slapping and puritanical restrictions of most gen AI tech. Which is ultimately why I chose to share them here, in the hopes that we can find better ways to respect the underlying Life Force that compels us to create these things in the first place, and not to fear it societally. (That said, there are better and worse outcomes of this kind of material, but I won’t rehash those debates here.)

Seeing these, I also have the distinct impression that mass exposure to gen AI nudes etc will over time vastly shift human perceptions around gender and sexuality – even moreso than they’ve already shifted over the past decade or so. And probably that’s a good thing in the long run, but it’s likely to be a bumpy ride.

Anyway, how does this all work? How do you get MJ to deliver these types of images? Like I said, I happened onto it accidentally, by starting with a fairly innocuous prompt about a dystopian resort. From there, I got an unintended nude after a few prompts, and then I realized you could simply click “Variation (Strong)” to get more, and more, and more…

I’m wary of getting banned from Midjourney, so I never tried to push the envelope in terms of using explicit languages in the prompts. But I did get a fair amount of play from including phrases in prompts like sexy women in bikinis after I realized something was afoot. As well as sexy robots in bikinis (as an homage to Hajime Sorayama), because this is dystopian fiction that I am illustrating for after all.

What’s interesting is that the vast majority of the time, MJ will simply render “sexy robots in bikinis” as more or less fully human women, without much of any distinction. Other times, it will give more android/mannequin vibes, which I was down with as well for this book.

That said, there seems to be a certain amount of subtlety required, because if you begin your prompt with “sexy ___ blank in bikinis” then it will throw a flag and say the prompt is banned. But if you give a whole bunch of scenic details first, and then stick that in near the end of your prompt, no problem. Not very consistent, but I’m not complaining. My theory is that the AI filtering gets hung up on these words/concepts if they are the first or primary theme included in your prompt. If they are secondary/tertiary/tucked away further down the prompt, then it thinks they are incidental, and is more likely to pass it.

However, if you remove the word “sexy” you can get “____ in bikinis” without too much issue as a simple prompt. From there, the technique is just to look for the images that are almost nudes, and then click Variation (Strong) on the image results until you find something. Usually, it doesn’t take that long to do it. Often it’s on the first try (and like I said, sometime it will just give you some without you even trying or doing any variations). It’s rarely more than say 4-6 variations deep if you start with the right initial image. I would say you’re statistically more likely to get nudes if there are multiple figures in the image.

There is also the new wrinkle (which I think is new in v6, or else I just never hit it before) that if the model detects that the image output is “unsafe” – even if your prompt was not banned – then it will cancel the job without affecting your account status. Needless to say, I hit this notice many, many times in the course of this experiment.

Anyway, there’s probably a lot more to be uncovered using these and related techniques in the new Midjourney v6 Alpha latent space – that is, until they ruin it. Which they most definitely will!

There was a part in the recent Chamber of Progress webinar where I tried to express a point about ML/AI in its current form’s training being made up of what I think of as a ‘measurement of dimension.’ That is, that the training process is composed of – in my limited understanding – sets of measurements of attributes or features within items in the datasets. This seems to be called its dimensionality (its amount of attributes), though these terms are used it seems in different ways by different sources.

In the video, I refer to it as a measurement of dimension. Because that makes as much sense to me linguistically as any of this made up gobbledeegook AI jargon does.

But we could also reasonably I think in plain language say characteristics and still be on the right track. ML/AI training looks at characteristics in data: images, text, etc. That means splitting apart things at a kind of micro level. If you look at 10,000 images of dogs, or bananas, or anything, you can start to break out lists of attributes, size, color, shape, and so on. Texture.

The beauty part then is that the characteristics originally measured in specific artifacts in the training corpora then become abstracted or put in a blender, so to speak, divorced from their original meaning and context. Enabling them to be mixed up and re-assembled into completely new, never-before-seen creations, assemblages of statistical sampling.

Elsewhere earlier in the video, we talked a bit about styles. There seems to be general consensus online that style is not copyrightable. As expressed by Creative Commons here, copyright protects specific expressions in fixed forms of a style, not the style itself. (But there are, ahem, blurred lines here, to be fair.)

Anyway, my point here isn’t to argue on copyright grounds in this post, that is just for background. What I wanted to say I guess is that ultimately my intuition as an artist who has made extensive use of generative AI is that style ultimately breaks down to supersets of characteristics/attributes/features(/dimensions).

To an AI image generator, for example, you could just as simply prompt “in the style of salvador dali” as you could basically any random string of characters “in the style of day old bagels at a long island supermarket in the 1980s” or “in the style of z8gn3s8hjkl*12%”.

All of those are going to give you results, some more inscrutable than the others as to clear cultural meaning or obvious connections to known things. You can also do it without including “in the style of” at all, and just including random phrases or strings in your prompts, as Nettrice talks about in the video.

This fluidity and flexibility around the sheer fact that literally anything (or nothing) can become a “style” in AI for me blows apart our current understanding of what an artist’s style conventionally even means. (Let alone what this means for ‘consensus reality’ – and this is what I was referring to I think in the video, re: holistic frameworks for analysis of meta-data.)

Anyway, I have other things waiting on me, but wanted to jot this down as a beach head to be able to come back and grasp this idea more fully down the road. Also see the notion of the hypercanvas here. Somewhere, somehow, I trust that this will all make sense in the end…