cry to the angels and let them swallow you

So, I've developed something of a thing for artificial intelligence machine learning. There's a lot of different ways to do it, and I've read articles on all of them, it feels like. But one that's growing in popularity is image generation. I started playing on Deep Dream, then, a few years later, moved to DALL-E (which is now https://www.craiyon.com/), then to NightCafe Studio, and finally I was accepted to the Midjourney open beta (that last link will open a Discord prompt to join their Discord server, which is how they enable the image generation system through a bot function).

Let's take an example I found through Midjourney. The search prompt was "beautiful woman modeling a stunning silk floral kimono outfit, in the style of Alexis Franklin and Tom Bagshaw, no blur". So this is what that turned up for Midjourney:



Great detail, wonderful differentiation, little fuzzy on the hat, and in the fourth image, the hand looks somewhat detached, but otherwise, this could be four different portraits hanging in a gallery somewhere.

Meanwhile, the same search term through NightCafe brings us, well...this:



Great pattern, style very similar to a kimono, black sticks for arms, headless, but decent.

And that brings us to Craiyon, where I ran the same prompt. Midjourney nets the user four to six detailed images, depending on detail of prompt, which the requestor can then reiterate, or evolve. NightCafe nets the user only one image, but generally extremely detailed, which can then be evolved. Craiyon does nine quick "sketches", and it's abysmal with facial recognition. These were the six that made the most visual sense:



Definite kimono, it's the right cut, the pattern's good, the pose is good. And at least with a side profile, it doesn't look so completely like the face is distorted.

And for all of these, I didn't chop off the heads, they came that way.



Remember what I said about their AI engine not quite..."getting"..faces? This is an excellent example. You can tell that it IS "face", but...not much beyond that.



The third only works, I think, because the head is cut off...



The fourth is not a kimono, but it's a great pattern, and it's a pretty perfect attempt at a sari. Face is still very odd.



This one's not as bad at some of the others, but it's still not the best face.



And this one could have walked straight out of a J-horror film.

There are some ethical concerns with the technology, to be fair--this article lays them out fairly well for Google's Imagen generator--but it's one step further for machine iteration in the goal for independent consciousness. And that is still fairly exciting to me.

Ultimately, we'll just have to realize we're moving into the era of post-reality, and adjust as we can.

here in the half-light, the motorhomes leave

This person does not exist. This is why. This is how you can help train the AI to create these images.

I find this both compelling and creepy. Compelling in that it's phenomenal we've come this far already in training artificial intelligences both to recognize what makes up a recognizable image of a human being, and allowing us as individual existing humans to help the machine learn. Creepy in that the first thing that crossed my mind was nefarious repurposing: how many of these images of nonexistent people are going to start showing up on Facebook profiles, Tindr and Grindr profiles, or used to 'catfish' other humans for emotional or financial grift reasons?

I've said this for a few years now, but this reinforces it: we now live in a time where we can no longer trust images, moving or still, to be true simply because we see them. This is further proof of that truth.

In other news, this wasn't surprising in the least. In fact, the only surprise I felt reading that article was that 25% of the grid being geared to adults strikes me as a tad low, and I thought it was higher. Also, the comments section for that post immediately went off the rails. I'm fairly sure I'm done commenting--I think I made my point, even if that point was instantly misunderstood--but I do find it interesting that the automatic first conclusion as to the ratio of sex/extreme content sims to non-sex/extreme content sims was 'dem filthy liberals'.