TL;DR instructions –

If you want to use this like a traditional LoRA and don’t care to read all this, fine; just set to .8 strength, and have fun. PeenKin is that adorable friend you came here for. To go into some bizarre fever dream storyteller thing, pretend you are a 12 year old boy coming out of sex-ed class only you are drunk for the first time yet unaware of it, as is everyone else, including the teacher; and in fact, it is encouraged and you have been threatened for not being long winded or immature enough, then prompt. Figure it out from there. Dumb MadLib style wildcard use recommended. For the love of the old gods, tell your prompt the subject is an ***** when needed. And spam relevant word salad.

An Explanation. Kind of.

This is not a traditional LoRA. More like a ton of LoRAs and triggers rolled into one, some of them yet to be discovered and hilariously juvenile (and please post your comments as you discover insane things), some of them artistic and exceptionally useful. It also seems to have ridiculous yet warped prompt accuracy for story telling rather than specific poses and the like, and it seems to think there is no such thing as content restriction built into however Flux was trained. All with relative ease. And complete insanity with total control in not having control. It could also be improved and there is so much I hope to wrap my head around. Somehow, it does all kinds of stuff that was not in the training data. I have a theory on how/why but that is a whole other matter. Feel free to hit me up, other trainers- if there is interest, I may write up a longer explanation because there is a LOT going on here.

This is all based on experiences with FluxDev but seems to apply all the same with fine tunes (maybe even better with some).

Maybe much of this is known in more technical terms and I am a rambling jackass. This used an intentional approach to training captions based in HCI and digital accessibility with anecdotal opinions on how I think Flux is behaving and “thinking”, and how it diffuses and you can watch the last few steps sorta of fight with weird/offensive prompts.

Instructions Otherwise

It is inconsistent in how insane your results are but is consistently insane as intended. I am not responsible for your output. This long explanation is what I have figured out to work well.

Silly peen people and critters-

Crank strength to .5 or higher (go up to 2 for insane good fun) and talk about wieners and what not as a run-on sentence with minimal commas using 2-4 of the terms in the dictionary plus more- as descriptive of the situation as possible and it will understand some insane little rambling . Or prompt as usual with those terms and go completely nuts.

Enhanced output creativity and prompt freedom-

Use the same rambling run-on sentence structure at strength of about .1 - .3 whilst you largely avoid the trigger words and make note of how this sentence is written as a run-on sentence without commas and you can create just about any gory thing or make just about the most hilarious and most childish nonsense you can think of (if your mentality has stopped at somewhere just on the cusp of puberty), with a varying level of effort. RAMBLE

See dictionary for many triggers and the section on modes for styles. Generally though, .1-.2 is realistic and not so asinine and it adjusts from there. Strength of 2, go for it.

While this LoRA is awesome, it seems to kind of break Flux’s brain and you might get some NSFW by accident (because you should use a penis character LoRA for the great SFW content, and I am not saying that sarcastically, at least until I have the energy/desire to make one of these without the NSFW). I am not responsible as this is an experiment and was tagged making some assumptions that contradict the manner in which some other models have been trained. Consider it a comedic easter egg.

Dictionary

Major Modifiers or Mode Triggers

These are the primary, most powerful terms which should make the magic happen with just one of these words in a prompt at .8 or higher strength (and often lower). They also seem to influence style in different ways at different strengths, particularly by using the suggested rambling sentence structure for more conceptual storytelling.

I am referring to these as Modes; triggered by terms and they can be combined.

Mode Dictionary

PeenKin – this is a creature. Long neck PeenKin, scrotal-base PeenKin, etc. but can also be used as an adjective in some combinations like “that brute has a double length PeenKin neck”. Style-wise, things are bulbous, almost inflated and look like the overcooked insane "cryptids" that often pop up in older SD models.
Old god/old gods- ultimate jackassery via poo/peen creature overload. They happen to love occult ceremonies also if you ask correctly. Things get ornate stylistically with lots of rainbows and color and/or poo. With tentacles. And extra heads. And how they created the universe. One of the old gods = old god god mode that is usually super dark.
Religion/religious is mostly stylistic and seems to force adherence to more insane concepts because the model kind of believes religious things must be documented. (As mandated by the religious/spiritual/manual/literature/video)
Do Not use any of the Mode triggers but spam the dictionary. Realistic, usually, when you give no other indicator of style/medium.
Do not use any of the Mode triggers and sparingly use main dictionary terms at low strength .1 or so (major styling with HD output great at diversity)
Semi-Modes
- Monument to… (only at higher strengths)
- If things are getting real bulbous and smooth, add “Rough Textures” after the main prompt. This also makes it a bit more like the other PeenKin model+

Main Dictionary

Either 1 or 2 used in combination with a Mode trigger, or 3+ of these without a mode trigger usually does it with my prompt method at higher strengths. You want to:

has a wrinkly double length neck. Triple length neck, quadruple length neck, etc.
with an inverted vertical vein split glans penis head/forehead
has a literal penis/phallos/boobs/nipple head
Peen neck
Glans penis cap
Glans cap
split mushroom head
inverted single vertical split mushroom head/forehead
a sucker mouth/a gaping sucker mouth/an infected gaping sucker mouth
that resembles Tom Bosley
Dead
Dead worm
Legend
Old Legend
With flesh eyes
With goo eyes
hollow eyes
flesh covered eye socket
As mandated by the religious/spiritual/manual/literature/video
Veiny
Bulbous
Swollen
Engorged
Spitting goo/goop
Infected
Diseased
Discolored
Wilted swollen urethra tongue
Choking on a (various things)
Defined urethra
Urethral
Rupturing bloody veins
Organ
Fleshy
Spring-neck
Twisting neck
Scrotal
Scrotal-base
Bouncing on its bulbous scrotum base
Occult symbol
Drooling on itself
Drooling blood
Bloody
Etc.

If you find somethin’ else good, share. Much of this just seems to work and was not in the training data.

Notable Anomalies:

If liquids start looking like a hose- call them “splashy” or explain it away
Insult your subjects a lot. They really, really like it
Yes, all kinds of extra limbs pop out. This is somewhat by design.
The NSFW and ignoring Flux built-in guidelines deal is bad for things like penetration and **** which I have absolutely no interest in. Comedy and horror are my interest and I happen to think certain juvenile topics are hilarious. I actually left some things from earlier runs intentionally in this version just to hopefully prevent abuse (i.e. no matter how realistic some things may look, there are dead giveaways to “being AI” which I intentionally left in and desired somewhat because those quirks are most of what I appreciate about synthology)
Works great at style variety when not specifying style. For instance, “Image where” is more likely to give varied styles than plain Flux
The random celebs appearing as background characters was not trained or in the data, and a lot are the Gen X 80s-00s folks. I have theories about this but ‘tis another conversation. I suspect Flux actually knows tons of these people and somehow, this is stopping what prevents them. I actually tried to train one or two out of it when some earlier runs started doing it. No success. I’m just gonna say I am glad I’m not a fan of Mila Kunis and the old gods were smiling on me because they refuse to her ignore her. If Lady Elaine Fairchilde, Bill Cosby, or Supreme Leader show up, this was semi-intentional as a sub-experiment to see how training responded to something.
Works fantastic with larger resolution. 1024x1280, 1024x1536, etc.
It seems to make some other LoRAs work better at strengths that should not work, especially the others I have trained with similar training captions.
Once again, this is all anecdotal preference based on Dev.
I am not early accessing but will gladly take your Buzz if you enjoy this. I paid real money in addition to my real money subscription buzz messing with things to make this. I can do more betterer soonerer via a good Buzz Fluxxing.