Monsters are real. After 6 years of Trump, I think we can agree on that. But not all monsters are Monsters. Some of them just want to move to Scottsdale, Arizona, buy a home in a luxury community¹, and send their kids to a small, liberal arts college in the Northeast.

This series, leading up to Halloween, is about that college.

Monday: Medusa’s Son

The Medusa’s son, nicknamed “Medi,” is a sophomore at this idealized, absolutely-not-stolen-from-a-Netflix-property college. Although he seems the party guy, he’s got a stable² boyfriend, a Faun he met at freshman orientation, and has barely let go of since then. Still, he’s not above finding some unsuspecting boy to join in the revels. He also loves swimming in the creek, and taking part in off-beat campus celebrations that allow him to be naked.

Medi and the Faun went to New York City for fall break. Medi really loved the giant mirrors and the bathrobes.

Medi is going to be at the naked swimming hole every weekend until it ices over. (The snakes hate cold weather.)

Last day of classes each semester is marked by the midnight naked run. Meet at the library, run to the president’s house. Medi wants to find out if the first 10 get the invite to a private party at the president’s house.

Most nights, Medi is out at a club with the Faun (next image), but sometimes, he goes to bed early.

He loves Tangent, the club, but it’s hard for him to stay fully dressed.

(Full prompts for the images are here (“dark” images) and here (“light” images))

But, I discovered a quirk with PbC: for best results, you need to prompt two things: 1/ the trigger word (“p3n15”), and 2/ An actual mention of penis.

Here’s what happens:

Neither

We get a “safe” option - a pair of cut-off jeans. (In contrast, Homo Fidelis does not assume clothes, and generates a nude)

Penis Prompt, but no trigger word

This is the most disturbing image: testicles are generated, but the shaft is not there. (Although, the testicles are called out specifically in the prompt)

Trigger, but no penis prompt

With the trigger word, but no explicit mention of a penis, Krea2’s native censoring likely takes over, and we get cut-offs once more².

And, finally, it works…

This takes a clear “belt and suspenders” approach to prompting: there is the trigger word as well as an explicit description of the penis(es) at hand. Then, you get good-looking, realistic penises in Krea2.

This was my first generation for this purpose - it’s also got DickTator and Penis Size (at weight 3 or 4). It’s got everything I kinda expected: DickTator’s habit of smearing everyone with cum, and Penis Size’s habit of making it too big. Other than that, pose / expression is good, color is good.

Second time, I wanted a full-length image, so I added prompt for shoes and socks. I got that, and also still got DickTator spooge. I like DickTator - it’s an all-purpose “I want male nudity without a lot of work” LoRA, other than that habit.

I wanted it to be a crowded club, but thisa wasn’t quite it. At this point, I had switched to Coachbate’s Penis(PbC), and added in Cinematic Dim. PbC definitely helped on DickTator Spooge, but not convinced yet that Cinematic Dim is bringing much to the party. (Also, Rob’s hand in the guy’s mouth is a bit odd.)

I tried adding in smoke, but it’s not quite what I wanted. There is probably a LoRA for this that I should find. But, there is now enough of a background that I can use other tools.

This is a pretty heavy-handed use of Lightroom to do some additional effects. If this were a “real” piece, I’d tone it down a bit so that Rob doesn’t quite pop against the background. I’ve been using Lightroom a lot more to do the final edits on an image, and it’s worth it. It was far easier to use it to make Rob a little brighter; I don’t even know how to prompt that.

What this means is that I’m taking the exact final prompt from the F2K9B version, and running it Krea 2 Turbo. The Flux 2 had no additional “Penis Helper” LoRA, while K2 had [Dicktator][] at weight .6. These were run on DrawThings; additional generation details for Flux¹ and Krea² in the foot notes.

The Basketball Game

Neither generated a basketball for me. That’s understandable, but a little odd. F2K9B struggled with dick, which will be a recurring theme here. The bigger concern is that the poses are awkward to impossible, and there are some random extra limbs and hands hanging out. The paint lines on the floor also make no sense. Almost all of this is solved in K2, and I think the proportions are slightly better.

Round: K2

The Shower Room

Once more, F2K9B gives us a nub instead of a dick; Only with the General Penis helper did it create a dick. But, K2 also had a helper as well. On the F2K9B, the floor seems to be some sort of [quasi-crystal][] (prompted as “hexagonal tiles”), and the shower heads were purchased at a flea market. There is also the general sense of claustrophobia, with the walls just a little too close, and the room a little too small. However, the soap and water is believeable. I don’t know if this is the LoRA or the model, but Omar is tall here.

On the K2 side, I think the lighting is more interesting (but not necessarily “better”). The tile work is really exceptional³. But where this falls down is the water and soap. It looks like it modeled the soap “assuming” the surface tension of water. But that’s wrong; one of soap’s key properties is that it breaks water’s surface tension, and thus, the image seems off. Also, once more, the room is a little too small.

Round: K2, but it’s not as clear.

The Locker Room

This is the scene where K2 seemed to stumble. Although technically perfect, both of the K2 images lacked a warmth that F2K9B seemed to generate in F2K9B. The F2K9B are lived in, and don’t have the same clinical perfection that the K2 seem to have. I do want to play around with this; it may be as simple as a pass in Photoshop to warm things up, and add a light haze.

This is not to say F2K9B is perfect - in particular, in two of the three examples, there are serious issues with hands. And in one, Mr. Nubby shows up again. But these can be generally addressed with a bit of in-painting afterwards.

Round: F2K9B

The Cafe

The result here is, again, not as clear as the earlier examples. It took me a moment to notice what was wrong with the K2 version - Omar and Kahlil aren’t looking at each other. This is likely a prompting issue, not anything with the LoRA or model. I’m not going to call this for either side; this is the most undeveloped prompt, so I’m not going to count my own mistakes against the system.

Round: Draw

Conclusion

Overall, I like K2 better than F2K9B - the images are better, if perhaps a little more clinical. In terms of raw numbers, there were 2 for K2, 1 for F2K9B, and one draw. I think that counts as a win for K2. At least, it was for me - it convinced me that for this particular chapbook, I’m should go back and re-create the images in K2 for the final version.

“…you say your interested in content but now interested in illustrations or paintings like the fuck”

This is an amazing quote, insofar as it creates a very specific definition of “content” that is very narrow and reductive: video, maybe photographs. The idea that there is something else is beyond her imagination.

Now, I know that there are terms of art, and here, “content” is a term of art. What bothers me here is not the narrow definition of “content,” but that this becomes a narrowness not just of medium, but also message. Online, browsing images and videos, there is a bland sameness to all the images.

Some of this is genre: mirror selfies impose location, framing, and lighting restrictions³. But some of it also seems to come from codification of fetish. That is, certain markers become not just a synecdoche for the fetish, but become the fetish as a whole.

The problem there is that these markers become gatekeepers. If the marker isn’t entirely someone’s thing, they are effectively kept out of the fetish. In addition, it punishes creativity and innovation within fetishes⁴. Online, a flash of personality seems rare: everyone sticks with the established tropes and images.

Part of my goal at [SpunPhotons][1] to to upset these tropes. [Monday][1] [Regendering][2] and [Random][1] [Remix][2] are my main efforts, but nearly every image I do, I try to do something with 15° of yaw in it.

There are some problems here. First, it was not written in typical narrative Flux style, but instead it was in what I call a “semi- Danbooru2” idiom¹, that is, a bunch of short concepts strung together. The nice thing about this approach is that I don’t have to do anything to distill the concepts: each comma-delimited clause is a concept.

This was 17 concepts. Listed below with “+” for present in the image, “=” for partial credit, and “-“ for not present

• = portrait photo of a muscular man sitting flat on floor legs apart,²
• • big crotch bulge
• • one hand behind his back supporting his weight,
• = he is wearing very short striped pink loose boxer shorts,³
• • massive legs
• = sweaty, ⁴
• • tilted angle,
• • dutch angle,
• • shallow depth of field,
• • extreme chromatic aberration,
• • volumetric fog,
• • neon lights,
• = cinematic lighting, ⁵
• • bright atmosphere,
• • film grain,
• = high-resolution realistic detail,⁶
• • charcoal grey gradient background,

In the end, this is not a great showing. Less than 20% of the concepts were definitively rendered, and over half were not there (or could never really be there.)

Now, the challenge here is that picking a prompt at random on Civitai is in no way picking a high-quality prompt. The images on Civitai are ones that the users like, and may not even reflect what their original artistic vision was. We see this in the presence of keywords that have absolutely no reflection in the final image⁷, likely relics of earlier exploration, or using old prompts as foundations. This re-use, while completely understandable, makes computation of “intended adherence” much more difficult.

There is still a place for the occasional “__ ignores everything,” when I find a suitable example. But, as long as I’m pulling examples from the open internet, it’s clear that at best, I am only looking at median exmaples.

But what to make of prompt adherence more broadly? Right now, I think it’s actually a red-herring. The current image models are definitely good enough for one-off images, where you can adjust the art direction on the fly. But the low adherence makes it hard to use for longer graphical story-telling. You can see this in much of the slop out their: character consistency just isn’t there. (I am reminded of a recent Youtube video essay on the Pennsylvania Railroad, extensively illustrated with UP engines.) This makes it hard to use for anything more than two or three frames. Layer so-so character consistency on top of that, and it seems nearly impossible.

So, what is to be done? I’m still exploring this, but right now, I think the answer is pastiche: one image is a combination of frontier-model rendering, coupled with in-painting, and specialized editors (e.g. face swap). Any “real” piece of GenAI “art”⁸ is going to be a composite of several images with a lot of spot-fixes. Until we have really good brain-computer interfaces, it’s always going to be succesive refinement.

When I first started writing this, I was Flux.2 Klein is a good model: high quality images, in only 4 passes. That’s pretty amazing. In addition, the LoKR seems

But first, for today, I just want to illustrate the problem. Here’s a prompt I was working on:

The image is set in the main suite of a converted commercial jet. Where ten-abreast seats once were, Omid is sitting in a single seat that has been placed right on the airplane’s center line. Two cars could park on either side of the seat; the luxury is the space and privacy.

And from that prompt, I get the this image:

This is not spacious. It’s a cramped land of paranoia and insecurity. And once you see it, you see it everwhere.

In the next one, I’ll give a few more examples and try to explain why we get here and maybe a solution.