Or is it “amazing footage of ‘8k’s?”

Whichever it is, the quality is simply stunning. It’s like I’m really there; like I could pick up one of those solid gold ‘8k’s and take a bite out of it. 8k is filled with chocolate and rainbows, you know. That’s how come it works so good in the prompts.

According to the expert,

the only thing putting 8k in our prompts should do is maybe make our images shiny. This does not surprise me – the amount of 8k video and 33-megapixel stills that were included in Stable Diffusion’s training data is… none. Not a one. LAION 5B’s data is all, or at least mostly, 512px * 512px – so ¹/₂k. I just do not believe that Stable Diffusion has any idea what we are asking for when we ask for 8k, If it learned from us, I could see it maybe figuring out that since we always ask for it with other magic words, we must want it to do magic, too. But that’s not how any of the AI art models work. We could get a bunch of 8k footage and and train our own copy of SD with it, but that has not happened with the big Stable Diffusion. Maybe when SD 3 comes out. But I doubt it.

But I have been wrong, oh so wrong

in my hypotheses before. At one point, I thought that any prompt longer that 300 characters would break. (I was mostly wrong – it’s tokens you need to worry about) I thought that “in the style of Greg Rutkowski” could only make pictures of hot cg women look worse. (I was mostly right – though there is at least one notable exception.) I even thought once that Stable Diffusion was trying to punish us for “stealing” certain artists’ styles by making us look like pedos. (I was entirely out of my gourd, but that doesn’t make Ilya Kuvshinov any less dangerous) Let’s try out some Ks and see how wrong I am this time.

All of these images come from Dream by Wombo. They are almost 2k images – shy by just less than 90 pixels. So we shouldn’t see any resolution effects from any K higher than 2k, as if those were even a thing. But I’m gonna stop carping and start sciencing.

Our lab rat for today is Redhead Space Marine in a Corset. I chose her because, apparently, I have a pauldron fetish. I did not know that until I saw her shoulder armor, and then, whoo boy.

Methodology:

To test this scientifically(ish), I took each “resolution” from 2k up exponentially to 512k, and tossed “HD” and “UHD” into the mix so we would have a couple more values below 8k. I went all the way up to 512k because that is 1000 times the resolution most things were scanned into LAION 5B. Also, that seemed to me to be the point where things would be reducto’ed close enough to absurdum for my purposes.

I tested each resolution’s effect at the end of the prompt, where 8k usually resides, and as the very first thing in the prompt, where it would be the most powerful. I ran each resolution through Wombo’s Dream App three times, picking the best out of the four choices Dream gave me each time. then I chose the best of those three. So, best out of 12. Not exactly a huge sample size, but I’m not exactly a scientist.

for the first run, which is the top row of the results below, I used this prompt:

nebulae in the background
beautiful redhead space marine in a corset
hard rim light
dynamic composition dynamic poses
[their "resolution", e.g. HD, UHD, 2k, 4k, etc. Control leaves this blank]

The resulting outputs were pretty good, but in them, I got:

• a few images without my beloved pauldrons (automatic disqualification)
• many with crappy eyes (Often the final deciding factor between the 3 runs was whether she even had recognizable eyes or not.)
• several with substandard décolletage (which I allowed, but only grudgingly)

If you look at the first row, you can see several images with a distinct lack of décolletage, or “cleavage window,” some have eyes that were the best I got, but not good at all, and although they did all get shoulder armor of some sort, a few of their pauldrons are vestigial at best. Barely even spaulders. So for the second prompt, I decided to risk throwing science out the window and change more than just the prompt order. Call me crazy, call me unscientific, but good pauldrons are just that important to me.

The second row is the second, improved prompt generation. Much better pauldrons, with several going to the effort of trying to show them off. In all cases but two, the décolletage is greatly improved, as well. And I don’t absolutely hate any of their eyes.

The second row is:

[their "resolution"]
space marine redhead in pauldrons and lace corset 
beautiful eyes
hard rim light uplight
nebulae in the background
dynamic composition dynamic poses

I’m putting my money on 4k. I think that’s the only k that Stable Diffusion could know, apart from possibly 64k.

Control

HD

UHD

---

---

---

---

---

---

16k

32k

64k

---

---

---

---

---

---

Analysis:

Let’s go through them k by k, k?

0k (CONTROL)

In her natural state, Redhead Space Marine is bold and assertive. She isn’t afraid to interpret orders in a way that sends a subtle “frak you” to the higher-ups. If you tell her you don’t like the way she’s looking at you, she’ll just put on sunglasses.

HD

She’s the newbie, and it shows. She tries to put on a brave face, but every once in a while, you can see the cracks in her courage. Row two has non-regulation pauldrons. Most make-up of any resolution, which Dream usually associates with “beautiful.” Technically, this counts as an increase in quality over Control.

UHD

Also has plenty of “beautiful” makeup, but has Control’s assertiveness as well. Her splayed arms say “What? I can’t wear makeup and kill alien scum? About the same increase in quality as HD.

2k

Quickest trigger finger of the bunch. Row two’s interpretation of “uplight” is hilarious. I honestly can’t tell if this looks better than Control.

4k

This is where we should be getting a real increase in image quality, but I don’t see it. She’s definitely more feisty than the rest. Row one thinks she can kick your ass anytime; row two knows she can. Still, I don’t think that is what most people who put 4k in their prompts are looking for.

8k

That goes double for 8k. Not the feistiness – this one is more on the sassy side, especially in row two. But I just don’t see an increase in image quality. Maybe we need to go higher.

16k

This one is the stone-cold killer of the group. She shows up in full armor for the first picture, and is the only one in the second row to completely refuse the “lace corset” order. (Though 2k has an interesting take on the order as well) I’d dock her points for no décolletage, but I’m too afraid she’d kick my ass. No increase in image quality.

32k

Not afraid to show her femininity. Row two says “You want décolletage? You can’t handle the décolletage!” Massive increase in décolletage quality, Negligible increase in quality otherwise, though.

64k

I was secretly hoping for Nintendo; I got horrific eyes after horrific eyes. Definitely no increase in image quality.

128k

Leather and lace. If we were judging quality on who had the best corset game, this one would win. But no increase in image quality.

256k

Doubles down on 32k’s décolletage. No increase in anything but cleavage.

512k

I was wrong. We hit reducto ad absurdum at least by 64k.

Results:

With the maybe possible exception of HD and UHD, none of these look measurably better or worse than any other. 8k definitely doesn’t look better.

(And poor ocularly challenged 64k looks so much worse). Even moving the resolution from the end to the beginning doesn’t seem to have an effect. The second row definitely has better shoulder armor and eyes, and in most cases more cleavage, but other than that I don’t see a huge improvement. Still, I stand by my 4k prediction. Maybe the image quality didn’t increase, but she embodies the Redhead Space Marine’s ability to kick your ass in a corset for sure.

If I wanted to be super scientific about it, I could run a few hundred tests of 8k vs control, but I’m satisfied with my results. And I just hit 8 minutes of reading time, which seems perfect for an 8k essay.

Do you know something about 8k that I don’t?

(That seems likely.) Was my methodology flawed? (That seems almost certain.) Are you going to keep using 8k because you are absolutely sure despite the evidence that 8k (or whatever k you have decided has the most magic) is improving your prompts? (Hey, you do you. Maybe the Placebo Effect works on Stable Diffusion, too.) Let me know the error of my ways in the comments.

Also, these results may only apply to Dream / Wombot / Stable Diffusion. You folk on Midjourney, DALL-E, etc. might get a huge effect from 8k. But I doubt it. If I’m wrong, please show me some examples in the most gloating fashion you can muster.

Pages: 1 2