Trying something a little more complex this time.
This was trained on 31 videos and 5 HD images.
Various phrases you can try:
{is kneeling in front of a man|is lying on her back in front of a man}(Mix and match as desired):
Cum shoots out of the man's penis and lands {on her face|in her hair|in her mouth|on her tongue|on her chest|on her pussy|on her eye}The man is stroking his penis back and forth with his hand(You can add in the handjob lora at weight of around 0.4 to improve the success of this one):
kneeling in front of a man and stroking his penis back and forth with both handsThe view is from {above|the side|POV from above}.Her mouth is {open|closed|open with her tongue sticking out}Her eyes are {open|closed}(Somewhat low success rate on this):
The woman flinches and moves away quickly in surprisemultiple people. Two nude women are kneeling in front of a manV1.1 Update:
I trimmed the dataset a bit to more evenly distribute the various concepts. The previous version had a somewhat low res look to it because I was using screen grabs from videos to get the HD images, and that probably introduced video artifacts. I got much higher quality photographs for the HD images this time.
I also trained on a lower learning rate for many more steps.
I think this version looks cleaner and seems to work better with character loras.
Description
FAQ
Comments (22)
What learning rate did you use? I've had issues with combining LoRAs if a motion LoRA is trained too quickly. If you used 1e-4, using 1e-5 for about 4k - 6k steps (assuming gradient accumulation of 4) might improve generalization.
Hmm, yeah that might explain it then. Unfortunately, this one took something like 8 hours to train as-is, so I'm not sure if I'll be up for some crazy long training at a lower rate.
@dtwr434 How many frames per clip? Did you use lr 2e-4?
@Dobo_J 45 frames, and yeah 2e-4.
What was your captioning style?
The example prompts are very close to how things were captioned.
Hello, you indicate trained on videos, do you have a tutorial or a link to carry out this type of training via video? I am currently training with image, I would like to test by video and I cannot find any link or tutorial that mentions this, thank you in advance for your feedback
If you look up my missionary lora, I included the training data, so you can see config settings I use and things like that. Just a heads up, the learning rate I use may be higher than the ideal amount, depending on how much time you're willing to wait for training.
@dtwr434 ok, thanks I'll look into that :)
@dtwr434 which repo did you use on github?
@co773c710n5 https://github.com/tdrussell/diffusion-pipe
@dtwr434 i have a lora trained, i do not use comfyui, i'm on an ssh gpu server, how do i start generating videos with it now? could you please give me the cli commands? i sshed into a gpu server, so it's not on my localmachine.
how many epochs?
@dtwr434 i created a new repo with diffpipe, check it out pls and try to generate a video using this lora : https://civitai.com/models/1100325?modelVersionId=1235997
@co773c710n5 I tried it out and had to change the example prompt a bit to get what I assume you're going for. I posted some example images. Nice work.
Can you please make 2 versions: 1 in which the guy jerks off and cum on the girl, and 2 in which the girl herself jerks off the guy and he cum on her.
As I understand in this version dataset is mixed?
Have you a workflow for testing hunyuan with loras?
When I try to use it, it's very pixelated but the ones submitted by other users aren't. What technique is used to reduce the pixilation. This only happens for me with this lora btw.
Are you combining it with another lora? If so, you need to use the block edit node. You can see an example of this included my blowjob lora. If not, I have no idea.
Actually, I think I can kinda see what you're talking about. This was one of the early ones I did and I'm guessing there's something wrong with the training data/method. I will try to revisit it.
@dtwr434 Thanks very much for the response. Very much appreciated.
@RandomPeach you could also try this https://github.com/facok/ComfyUI-HunyuanVideoMultiLora