In The Night Cafe

In the "AI Used In This Post" footnote I include with every post that uses any form of generative AI (Well, when I remember...) I almost always mention NightCafe. I can't remember when I first found the website but it must have been at least a couple of years ago and since then I've rarely gone anywhere else for my AI image needs.

AAM XL Anime Mix v1

Until today, when I did a bit of googling in support of this post, I had no idea NightCafe was an Australian operation. I also didn't know it had a mission statement: "to democratise art creation". The site's "About" page is worth a read. I won't paraphrase it here because it's a two-minute read but I recommend a click-through.  

I'm currently reading Supremacy by Parmy Olsen, which tells the story behind the development of the biggest of the Western AI developers, Deep Mind and OpenAI. I'm not far into the book yet (It's my bathroom book so I'm reading it a few pages at a time.) but the most surprising thing to me so far is the extent to which the individuals who started all of this had existential, humanitarian and even altruistic reasons for doing so.

I believe part of Parmy Olsen's purpose in writing the book was to show how those lofty goals became subverted by capitalism but I won't pre-empt the twist ending. I mention it mostly as a corollary to the brief description of Angus Russell's flash of inspiration for NightCafe, which was when a friend came round to his apartment and commented on how bare it looked with no pictures on the wall.

Nightmare Shaper v3

Somehow that led Angus to start a crusade to "allow anyone - regardless of skill level - to experience the satisfaction, the therapy, the rush of creating incredible, unique art." I'll pass over the inevitable argument that kind of statement is bound to start because this - believe it or not - isn't a post about the rights or wrongs of generative AI.

No, this is just another gamer's blog post where someone tells you how they've been grinding and now they've leveled up. Or something. 

Since I started using it, NightCafe has become increasingly gamified. There was always a log-in daily where you got five credits just for visiting the site each day, and there have been competitions running for as long as I can remember, none of which I have paid even the slightest passing attention.

A few weeks ago the administrators (Probably Angus's wife, who's the  Chief Operating Officer.) decided to turn that daily log-in into something a little more active, so now you have to make a picture before you get your free credits. It's a good deal because the minimum payout is still five creds but now there's a dice-roll to see how much more you get on top of that and on a good day you might get ten or fifteen.

DALL-E 3

Better yet, there are streaks. Every five days gets you a bonus ten credits and after a certain number of days you get an extra bonus and a title. I'm an Owl, which comes at thirty days. Next comes Horse at a hundred, then Bear at two hundred and Eagle after a full year.

I'd be further along but I missed a day because for some reason, one day I didn't get the email reminder. Until then, it had never failed to appear in all the time I'd been using the site. Now I take great care to check in every morning, email or no, because the penalty for missing a day is to slide all the way back down the snake to the start.

Not that it matters. The credits keep rolling in just the same and what difference does it make whether you're an Owl or an Eagle? Except it will matter to someone because there's a Leaderboard. 

I didn't know until today there was such a thing. Now I do, I still find it hard to believe. Who makes a leaderboard for a utility website? What would be going in someone's head to do a thing like that? 

Ideogram V3 Turbo

Top of the table today is Amelezz with a "1.2k day" streak, which suggests the streak thing has been going on a lot longer than I thought. If not for that one slip, I might be somewhere on that leaderboard. I doubt I missed many days since I first discovered the site, if any at all, and that has to be at least a couple of years ago..

But even then, my name still wouldn't show up on the board because I have never registered an account under any name. I must have filled in some form long ago or else how would they know where to send the emails but it doesn't seem to have required anything more than the address because my creations are all still accredited to "Anonymous User". 

Unlike Suno, which has a similar pattern of gamification and where I'm unknown as That Darn Cat, no-one is going to find anything I've done by searching Nightcafe. Mind you, they won't find much at Suno either, only the experiments I made when I first found the app, also a couple of years ago. All my recent work there remains resolutely private.

Google Imagen 4.0 Ultra

I am not tempted by leaderboards or likes and I am not a fan of gamification. In fact, I'm against it. When the new system was introduced I was irked. 

I could just have ignored it and stopped collecting the credits. I already had plenty, getting on for three thousand, and since I only use a handful every few weeks, whenever I need something for a post, I was hardly likely to run out.

I wasn't so angry I was going to do myself out of a free supply of credits just on principle, though, so I contented myself with some passive-aggressive push-back. I decided I wouldn't give NightCafe the satisfaction of seeing any new prompts from me to earn those credits. Instead, I'd re-use the exact same prompt every day.

For the first day of the new regime I pulled up one of the images I'd used for this post, which at the time was the most recent use I'd made of the app. I cut and pasted the prompt - "Walking through corn fields Covered in dust Lost in this dustbowl young female figure, old, worn clothing, line art, color, retro-futurism" - to generate a new image and that's the prompt I've been using every day since.

Fluently XL

The whole point of NightCafe, as opposed to other AI Image Generation sites, is that it collects together many, many models. The Pro version, for which you have to pay, has the most but even the free version has around sixty. 

At the start of my streak I just used whatever model I'd left in the chamber after the last time I visited but pretty quickly it occurred to me to try a different model each day. Had I been planning it, I'd have started at the top and worked down but as it is I've just been picking one on a whim and trying not to duplicate too much. At some point, I need to go back and note down all the ones I've used so I can make sure to try all the others.

It's been fascinating. The similarity of the images is striking but so are the differences. As of today I have forty-seven images from the same prompt which, you'll notice, does not mention anything about the way the figure should be facing, their ethnicity, the time of day or what exactly I mean by a "corn field".

All of these and many more are details left to the AIs to work out. It's very instructive to watch them doing it. 

Mysterious XL v4

In so far as it's possible to tell, all the women are white, often typically northern-European in appearance. There are one or two where it's slightly ambiguous but there are no dark-skinned faces and most of the clothing is broadly western in style. 

That seems telling because the prompt makes no mention of ethnicity or global location but although I didn't realize it when I came up with the prompt, "dustbowl" is a term apparently quite specific to 1930s America, as made familiar by movies and novels in general and Steinbeck in particular. I always imagined it to be a lot more generic than that.

That goes some way to explaining the clothing, too, which is vaguely appropriate to that setting with a lot of blue jeans, denim jackets, and long canvas dresses. Again, I only specified that it be "old" and "worn", something not all the models seemed to notice.

Neither did I say which way the figure should be facing but only nine feature a figure walking away from the viewer. All the rest are walking towards the camera except for one extreme outlier, apparently crossing the path from the left to the right. 

Juggernaut Flux Lightning

As for age, the various models took a fairly broad interpretation of "young", with the female figures
seeming to vary from mid-teens to mid-thirties. Two of the models ignored the instruction to make the figure "young" altogether and went for a white-haired, much older woman instead. 

All of the models, without exception, interpreted "corn field" to mean maize or sweet-corn., something I suspect shows a very strong North American bias in the data. When I wrote the prompt I wasn't imagining anything like that, as can be seen in the original short story from which the quote was taken. Where I come from, a "corn field" loosely means any arable crop. I was most likely thinking of wheat or barley. Certainly not maize. It also seems to clash somewhat with the aforementioned tight interpretation of "dustbowl".

Only a handful of models added anything significant to the background and when they did it was usually something that gave the image a somewhat post-apocalyptic or sci-fi feel. I'm not really sure where that comes from. Maybe it's just me, seeing things that aren't there.

Animagine XL v3

Another handful gave the figure something to carry, almost always luggage of some kind. There's a single image in which the woman appears to be holding some kind of gun. Most of the women are bare-headed but there's a smattering of hats. One very odd take puts the figure in what looks like a space-suit.

Given that I didn't give any suggestions at all on palette beyond the single word "color",  the collection is remarkably consistent, all shades of yellow and brown. Sepia is a particular favorite. Occasionally there's a pleasant blue sky. The very few images that use a lot of color really stand out. 

Every shot is taken in daylight, too, although a couple look like the sun is about to go down. Again, that was never specified or even mentioned so presumably "daytime" is some kind of default.

This degree of relative consistency is something I've also recently noticed as a feature of Suno, where I've had cause to reuse the exact same prompts multiple times, with just a single word changed, trying to fine-tune something. It's clear that whatever the mysterious black-box processes behind these images and sounds might be, they're far from random.

I'm very pleased with the portfolio I'm building up of these unnamed women, making their way across a somewhat forbidding landscape. I intend to carry on re-using the same prompt at least until I've run it through all the available free models, not so much to prove or test anything but just because I find it really entertaining to see what comes out each time I press the button.. 

When I'm finished, maybe I'll pick a dozen or so of the best ones and have them made into a poster. It could replace my Bojack Horseman by the Pool that faded so badly in the sun I had to take it down. Nightcafe does offer a print-and-frame service but I think I'll be doing this one myself.

 

Notes On AI Used In This Post:

Google Imagen 4.0

The prompt is in the post. All the models are given in the captions except for the four shots in the header image. The upper pair are both by Flux from before I started using a different model each time so it's no surprise they're so similar. The two beneath them are by two versions of the same model, HiDream I1, so again the similarity is explicable, although not so much the congruence of those two with the ones by Flux. The variant models for those two are HiDream I1 Dev on the left and HiDream I1 Full on the right. I notice the "full" model is a lot sharper and clearer.

My favorite, by some margin and not just of the selection you see here but of the entire run so far, is  the fourth from the top: the mid-teen girl in dungarees, carrying a heavy-looking sack and staring down at her feet, looking worried. That's very close to the image I had in my mind of the unnamed girl in the story and the very dusty, red-tinged field she's walking through is very like the setting. Perhaps surprisingly, that one's from a model one of whose whose biggest claims is that it "Excels in typography, producing text that closely resembles human-designed artwork."

I didn't use several of my other favorites because I was trying to give sense of the range and and to illustrate certain points I'd made. Before the event began, I was toying with the idea of posting one image a day throughout Blaugust and I might still do something like it, after it finishes. It just seems a shame not share the good ones.