This model depicts an male human masturbating in various positions (standing / sitting / kneeling) in a third person's view.
Promt example:
"In a crowded soccer stadium arena a nude male is jerking off in a seated position. He is sitting among the audience of a soccer game. He is grabbing and stroking his own penis using one hand, which moves back and forth rapidly. The audience conists of adults. The soccer game is visible in the background."
Usage:
0.8 strength (double blocks)
_________________________
Hi everyone, I love this community and I'm happy to give something back, although I don't have deep knowledge neither about training or the terminology used to describe it. This is my fist Lora and any kind of suggestions for improvement is appreciated, as I think I will continue my journey. So please help me out on the questions at the bottom of the text, if you could spare the time.
About This Lora: It is the 4th iteration:
3250 Steps 255 Epochs. Trained on a 4070 Ti Super (16gb) in about 8 hours.
The dataset consists of 15 Video clips: Each encoded to following properties:
512x512 / 24fps / 65 frames per clip
The dataset consist of men masturbating in different positions aged around 20 - 40.
They are mostly viewed from the side and front. No POV footage.
Most of the clips do not show the males face. There were 2 cowboy-shot / closeup clips in the mix.
The majority of the caption contains mainly:
"A nude male is jerking off in a kneeling/standing/seated position. He is standing/kneeling/sitting on a bed/couch/the floor. He is grabbing and stroking his own penis using one hand, which moves back and forth rapidly."
Trained with: musubi_Hunyuan_Lora_Tuner / toml file:
resolution = [256, 256]
batch_size = 1
enable_bucket = true
bucket_no_upscale = false
target_frames = [1, 25, 45]
frame_extraction = "head"
Beginners questions:
As I already read through a few guides, some are not really
speciffic on what learning rate is "too high", or how many epochs should be enough,
so I left everything as it was and kept adding epochs ot the weights This is what I used:
Is it appropriate for the dataset given above? What can help to improve the model /speed?
Learning Rate: 1e-4
Gradient Accumulation Steps: 4
Blocks to Swap: 20
Training Alpha (1-128): 16
Thanks to: The guide by ttplanet and the guidance of tdwr434:
https://civarchive.com/articles/10335/hunyuan-video-lora-trainning-with-gui-in-windows
https://civarchive.com/models/1087616/hunyuanvideo-pov-missionary
Description
FAQ
Comments (10)
Been waiting for this one, thank you for making a male focused LORA! Between it being male focus and a third person perspective, it’s exactly what I was looking for!
Thanks tons, got excited as soon as I saw this, it's great seeing males get more lora love, hope to see more.
Here's some initial feedback (all testing done at 320x432): Dunno if just me but definitely noticed major quality loss when using all blocks, so early on in testing I wound up using double blocks exclusively, I can't think of another lora I've had to do this on.
Additionally, I noticed jerking and hands got garbled any time both hands were near the penis (which I've also noticed on other loras) so I started working on creating a prompt to make only one hand be jerking and the other be to the side, so far this one has been noticeably better: "He is grabbing and stroking his own penis using only his right arm and right hand, which moves back and forth rapidly. His left arm and left hand is resting at the left side of his body." but perhaps we can find a prompt better than that as well.
Prior to this lora I was combining the futa masturbation lora and natural men lora to eventually get non-futa gens I liked (there's also the unretracted foreskin lora for anyone wanting to experiment with combining). But early on in testing this lora I found that combining it with the futa lora definitely improved the motion while also making cocks less huge than the futa lora skews towards and giving me better results I liked than either by themselves. Still testing what strengths I like most but it's looking like both under 1.
Sadly I don't have training advice since I haven't began yet, so can't say anything more than referencing the training info some lora creators have on their respective release pages or yt. But I can say this has already been useful and look forward to seeing it develop <3
Thanks for the comment.
Regarding the double hand motion I expreienced the same interference. It's strange considering, that the training material was clearly one-handed. Thanks for the prompt-advice.
I agree with the "double blocks" phenomenon. As I stated, I lack a lot of knowledge and I don't know hot to influcence the weighting System. I can confirm that the only decent results could only be achieved with 0.8 weights at double blocks, and I don't know why, or wheter it was just my set-up.
As the previous iterations were undercooked, I think that I might give this lora another round in the future.
Nice feedback!
This has got to be one of the funniest loras I've seen.
You weren't kidding! 🤣
You're a godsend.
Thank you!
How can I slow down the hand? Its often too fast. I tried to change my prompt but its mostly very fast.
I'm sorry I have no special promt that could address that, but I have a suggestion: You could always try Video Interpolation and Slow down the video that way. I know loras that are to slow and increased the video speed by increasing the framerate. You could do the opposite by increasing the frame count while keeping the framerate the same, if this makes sense...
@Adaptalab0r I am not yet deep enough in this topic but I will experiment a little bit. Thank you!
I wish people were allowed to do stuff like this in public.