This LoRA makes a penis grows longer and larger.
Version 0.2: can create a greater variety of angles. It work for males, females and even objects. It's a T2V Lora but it works with I2V as well.
T2V: The images from the 0.2 HIGH showcase gallery are all generated with T2V.
I recommend a strength of around 0.5 and a very descriptive and clinical prompt to avoid problems.
You can use it with the following keywords:
{Dorsal surface|Ventral surface|Side} view of {his|her} {erect|flaccid} penis growing from {small|average|big} to {big|long|girthy|veiny|huge|gigantic}, pointing {up|down|straight} relative to {his|her} body.
Example prompt: Video of a naked man in a bedroom. Ventral surface view of his erect penis growing from small to big and girthy, pointing up relative to his body.
It works with any prompt mentioning a penis, growing or not. Sadly, it struggles a lot with the orientation of the glans and the positioning of the penis if you are not descriptive enough. You can also resolve those issues by stacking this LoRA with an other penis LoRA. I had good results mixing it with Smooth Mix Futanari.
I2V: The images from the 0.2 LOW showcase gallery are all generated with I2V. The source images are not mines, they were taken from CivitAI.
The training data contains no faces so it works well with I2V. It arguably works even better because it doesn't have the issues with the inverted glans or attaching the penis to the wrong location.
I recommend a strength of around 0.5 as well for I2V.
Version 0.1: It can only really do one angle, character viewed from the front or facing the camera with the penis pointing down. It works for both males and female characters.
The trigger word is penis_growth but just mentioning a penis growing in your prompt should work.
It works best for me at medium-high strength.
HIGH: 0.6-0.9
LOW: 0.6-0.9
It seems to work for I2V also but I did minimal testing.
Description
v0.2 LOW noise
FAQ
Comments (9)
Interesting concept!
if you show me your captioning style maybe I can suggest some enhancements, the concept here seems well learned but the flexibility in the LORA seems low
I'm here to learn and I think I didn't pick an easy concept :)
The data was pretty much captioned like the prompt example but with the background added in. "Naked man on a white background, {Dorsal surface|Ventral surface|Side} view of {his|her} {erect|flaccid} penis growing from {small|average|big} to {big|long|girthy|veiny|huge|gigantic}, pointing {up|down|straight} relative to {his|her} body. Camera or character motion if any.
The data itself is just that, naked headless men on a white background. Most just standing, some are turning around, some of them the camera is moving.
The stuff about the surface view and the angle relative to the body was added after a failed attempt with similar issues.
I'm open to suggestions. There's a few things that I think caused the problem.
1- I flipped the training videos on both axis in an attempt to provide more angles. I knew this could be a big cause for problems but I did it, maybe foolishly, to try to increase versatility. This is the most likely culprit I think.
2- From my testing, Wan seem to have a strong understanding that small = flaccid, big = erect. I was thinking maybe removing erect and flaccid from the captions and just use a relative direction. Basically, I think I think I confused the model by calling those unrealistic poles, flaccid. It was fine in v0.1 of the LoRA because it was a way simpler concept to learn.
3- I used the default caption dropout, 0.05 I think. I wanted to try without it.
4- I didn't use regularization images.
I also left this one train for way too long because I kept thinking that it was finally learning correct anatomy. It was most likely just lucky samples. I fell to the sunk cost fallacy so it's way overtrained and I will probably never train another one as long as this one. That was a costly mistake. That's also why I recommend using it at 0.5.
I released it despite it's problems because it can do some pretty funny stuff like the flower pot and the trees in the samples. It also work extremely well in I2V in my opinion. It definitely understand anatomy better with a starting example.
@qcgooner Nice write up! if you're interested in me rewriting one of your captions, maybe upload just one clip and one caption and I'll modify it and send it back!
In terms of the lack of flexibility, I noticed it can only do white men, but yeah that would be caused by overtraining for sure, and maybe not just a lack of the word Caucasian...
Thank you for the help. I appreciate it. I created a zip file with a few samples of the training data and I added it to the 0.2 HIGH model. It's a good representative sample. Other videos are similar with slight variation.
As for the ethnicity problem, I was able to change ethnicity earlier in training but I admit I forgot to test this on the final version. Perhaps lowering the strength can help with that. The LoRA is still pretty strong at 0.5, I successfully used as low as 0.3. Ethnicity is definitely something I will add to the captions in the next version.
I also plan to redo or refine the training data for the next version. The training data was generated when we didn't have many wan LoRAs. I used Pony pictures and animated them with first frame last frame in Wan 2.2. I want to remake them in the same style but better quality with what we have now and what I learned.
@qcgooner Cool, so a couple notes, all for educational purposes, and this is just my two cents I'm not claiming this is all 100% correct, who knows! lol
With that said, here is my rewrite of the first file front-6
A Caucasian man's penis becomes larger,
His penis is flaccid, hanging down and initially average-sized, it grows larger, veins bulging until it becomes very large, pointing down relative to his body
bright lighting, white background, medium close-up view
Versus yours, I:
a) start with a key phrase "penis becomes larger" or "penis grows" should be in the first line and unchanged in every single caption. "A Caucasion man's" is a prefix and will probably in this case be rather static as well.
b) You use the word "flaccid" three times. I'd be careful with that, it might be overdoing it, and flaccid + large = not logical, so having that once referring to the starting state only will probably help. The base model is going to be trying to assign flaccid to all frames here, because it shows up so much, but that's not what you want, I think.
c) "Dorsal surface"... I don't think that means what you think that means. Dorsal means 'back'... no back here...
d) Occlusion: "The grown penis obstructs the view of his testicles commpletely." WAN doesn't do occlusion well, we don't see tentacles, don't mention them :) Or flip it if they do "testicles partially visible" instead of "anything" hidden.
e) You use the word small but unless you're rather gifted, that dick starts out big... I'd avoid the word small for footage like that, it's inaccurate.
My rewrite of the second one:
A Caucasian man's penis becomes larger,
His penis is small and flaccid, pointing up relative to his body, it grows larger, veins bulging until it becomes very large,
bright lighting, white background, medium close-up view
@az420 Thanks for the input. I will definitely try your style of captioning in my next attempt. I can see why it would work better. Starting with the phrase describing the concept first makes total sense, kinda like a trigger phrase.
"bright lighting, white background, medium close-up view." This should the background from baking in more and maybe fix the bad coloring a little bit, I like it. Adding the ethnicity makes total sense as well of course.
Some stuff you mentioned was added after my first failed test run which had pretty much the same orientation issue and it was unable to make them pointing down at all. I added flaccid to the end state because of that but I agree with you it shouldn't have to do it and it could even confuse the model. I'm starting to think the term flaccid alone is confusing the the model.
Dorsal surface is the right medical term for what the video shows I'm pretty sure but I'm not a doctor. I'm not sure it's a terminology wan understands anyway. It was also added as an attempt to fix the same problem and I don't like having that it in the caption anyway.
For the object occlusion, it makes total sense. Thankfully, I think the video in the sample is one of the very few where I described it like that. In the majority of videos they are fully or partially visible and described. I will keep that in mind.
For the relative size, my reasoning was to caption as small the smallest ones in the data.. Not necessarily what i subjectively think is big or small. There isn't that many true small ones in the data. My reasoning was that since wan is censored, the smallest ones I am showing it are the smallest the model will know about. I don't know if that makes any sense.
I'm also more certain than ever that I shouldn't have flipped the videos during training. Especially the Y axis flip. I think it would work only if I flip them manually and caption them accordingly. Something like "A man upside down". It's probably still confusing to the model.
I wanted the LoRA to be able to do funny stuff like a penis growing from the ceiling or the wall or anything else so that's why I tried the Y axis flip but I'm pretty sure it was a bad idea now.
Thanks for the input again, I will definitely try your recommendations for next version.
@qcgooner happy to share strategies, hopefully it helps :)
Does it work when the penis is... well, already penetrating something? 😅 Thanks 🙌
I tried mixing it with other LoRAs and it's different for each of them. It work with a few of them but you will rarely get the movement from both LoRAs at the same time. I find it hard to mix motion LoRAs together in general. It can be done but you need to find the perfect mix of strength for each. You can get very weird results with the wrong mix.