Disclaimer: This is my first time training anything so there are likely some really amateur mistakes.
Info
I'm posting this as sort of a sub-alpha mad scientist experiment. I'm also including the training sets (which were generated using Chroma Krea2 and Wan 2.2 / 2.1 with some of my fav penis and cum loras) so if anyone wants to poke around under the hood or make their own version? Go for it! ^_^
This latest iteration (-0.02) is a cleanup and re-run of the same dataset only for MiniMax. It basically fixes how the default model handles existing peeners and spooge while giving it the ability to add a peener without making a horridly mutated penis like thing appear.
The basic idea behind this is to see if LTX (and now MiniMax) can better handle longer generations with lots of movement without mutating a male subject's peener. That and honestly I miss being able to generate some epic hyper-cum shots I was able to get in previous versions of Wan.
Usage:
v-0.02:
Made a whole bunch of cleanups on this one. Basically fixed up and rescaled the original video samples that define the movement and spooge (oops, they were at 16fps), redid all of the references images using Krea2 in a higher resolution with more detail, and unscrambled a whole lotta messy captioning.
This version also adds a few first person POV shots from the subject's perspective looking down at his peener. This seems to work as intended for first person shots with existing peens, but if it has to generate one it tends to come out backwards (oops - probably more segmentation needed in the captioning).
Again, this one has same usage as prior versions, only for MiniMax! Trigger word is... PEENER which seems to be optional now.
v-0.03:
Same usage as prior versions, only for MiniMax! Trigger word is... PEENER although that seems optional in this go-around.
8/24/26 - Added a slightly less over-baked version at step 4000. This release seems to have an issue where the captioning managed to bake some scene elements from the source images in with the peener concept, which seems less bad at this iteration.
Note: If your subject is clothed and you want them reveal their peener without having it poking through their cloths, don't mention anything about their penis, erection, etc. until the reveal part. For example:
Orion is a 20yo male wearing basketball shorts leaning against a wall. He looks at the camera and shouts, "SURPRISE!" He hooks his fingers into the waistband of his basketball shorts then pushes them down to reveal his big erect penis.
v-0.002:
Trigger word is... PEENER
Prompt segments to make stuff happen are...
"with an erect penis" - to indicate the subject has a penis.
"He reaches for and grasps his penis." - to grab his peener.
"He's grasping his penis and he moves his hand rapidly up and down the shaft of his penis." - for wanking.
For normal spooge: (note "sticky liquid" was changed to "cum":
"He's having an intense orgasm."
"He ejaculates white sticky cum from his penis."
"Cum is squirting from the tip of his penis."
For hyper spooge:
Add in the words hyperspermia or hyperspermic.
"He's having an intense orgasm."
"He ejaculates a flood of white sticky cum from his penis."
"Cum is gushing from the tip of his penis."
For common spooge landing points actions:
"Cum is splattering on his legs."
"Cum is splattering on his abdomen."
"Cum is dripping from the tip of his penis."
"Cum is flowing down the shaft of his penis."
v-0.0001:
Trigger word is... PEENER
Prompt segments to make stuff happen are...
"with an erect penis" - to indicate the subject has a penis.
"He reaches for and grasps his penis." - to grab his peener.
"He's grasping his penis and he moves his hand rapidly up and down the shaft of his penis." - for wanking.
"He ejaculates white sticky liquid from his penis." - for normal spooge.
"a huge flood of thick white sticky liquid gushes from the tip of his penis." - for hyper spooge.
or possibly adding in the words hyperspermia or hyperspermic.
Meta Ramblings:
As I'm running this through its paces I'm already forming a to do list of things that might need tweaking or improving, so as time permits I'll likely post more updated versions. I'm also open to tips and suggestions, so pile em on! :D
So, queue the mad scientist experiment! I started off by generating a bunch of static images of peeners in various poses and angles. That alone seemed to give it an idea of what a penis was but it tended to stretch and distort when any sort of movement was prompted in.
After that, I tried mixing in some video clips from the internet which seemed to help But I quickly discovered what a pain it is slicing and dicing and editing those into a format the trainer could use.
So in the end I scrapped those and wound up just generating a bunch of short clips with Wan at various angles and common movements like wanking, walking, jumping, sitting, etc. And for the added icing added in some normal and hyper spooge clips in various poses both wanking and hands free.
Description
Retrained the same dataset used for the LTX version on MiniMax.
FAQ
Comments (16)
I salute your work, sir!
Why thank you! ^_^
I don't want to be too harsh since you're apparently new to training, but "the training sets (which were generated using Chroma and Wan 2.2 / 2.1 with some of my fav penis and cum loras)"
FFS... WHY would you train a superior model on outputs from inferior models??? Please just use real videos. You're baking in artifacts and limitations of those models and loras.
Well, I did try that but most of the real world videos I could find of hyper cum and penises were pretty low quality, in odd angles, or had garbage overlays from porn / cam sites that caused the training to come out like crap anyway. Truth be told, this is giving me mostly what I was looking for with some random annoyances like disappearing backgrounds and unexpected camera cuts. But in all I'm kind of getting burned out on all this AI stuff anyway so I'm probably just going to abandon this.
it's fine using synthetic data to train models.
'You're baking in artifacts and limitations of those models and loras.'
Almost all LLMs use synthetic data, even Krea2 and ID4 used synthetic data.
@wiiz I looked through the dataset and yeah, I can see some spots that should be cleaned up or redone. In particular the static images. Btw, I did include the dataset I came up with as a download so you can see first hand what's in there. And my horrid captioning. >_<
@wiiz The entire concept being synthetic data is completely different to having some synthetic data mixed in. It also makes no sense to do when source data that isn't synthetic is available. How do you think the original loras were trained? It's like training on 256 color gifs instead of the source HD video.
Furthermore, outputs from worse models is not the same thing as generic synthetic data, like 3D renderings, which obviously have to be synthetic.
@wiiz You should be aware, for next time.
creators of Krea 2 explicitly stated that they used no AI-generated images in their pretraining mix.
In their official technical documentation, Krea AI explained that they designed custom in-house classifiers to filter out synthetic data, noting that even a small proportion of AI images introduces unwanted biases and sets an upper bound on quality.
The white fluid on the tip is an unrealistic deal breaker, totally breaks immersion. I admire the effort though.I trust you see how the verisimilitude is off.
Thanks for the input and noted for the next go-around. It seems to do that at random, and I /think/ I spotted where its coming from in the sample videos so it should be easy to chop out.
@OrionMoonstone A fix would be awesome. Keep up the great effort! I know these things are finicky and time-expensive.
@ltcdrbroccoli Most likely this weekend when I can recharge my RunPod and I have some more time to mess around. I also really want to fix that backwards penis problem with the generated POV shots. >_<
@OrionMoonstone respect man. Thanks from both heads.
@ltcdrbroccoli Wait it just dawned on me... Are you secretly a Klingon? Having two heads and all... ;)
@OrionMoonstone qapla'!
@ltcdrbroccoli Rawr! >:D