Wan (hucows.com version): This model was trained on 20 videos from hucows.com as a test to see how well it can learn the motions from the videos. High chance for logos as i was not willing to erase the logo fom 6000 images.
Flux v2 and Wan v1: Same dataset as Wan version. Difference to old version, its trained on more variete of data instead of 90% realisitc its now about 50% realistic, 25/ cartoon/anime and 25% 3d cgi. therefore this version is not as restricted to stationary restriction and can place milking machines much more freely while still being able to maintain the old look. also its retrained to capture more data.
SDXL/PONY/FLUX V1: These models were trained on a collection of 150 images and work well for realistic as well for 2d images.
Description
FAQ
Comments (11)
You trained this LoRA solely on still images? Interesting. I didn't know it was all that effective to do that. So it is possible to train a video LoRA if you have enough still images to get the concept down?
That's generally how any lora works, since SD1. If you pick the right photos at the right angles with the right details, you won't need many. The complication comes when you're trying to do something the model has absolutely no idea about. However, the model knows women, might even know breast pumps and restraints. So it's not too many jumps.
However you'll see even from the example videos, there isn't a huge amount of movement. I'd assume this is what you'll see more often with only image based training and with high strengths -- it'll abide to all it knows about the movement.
Which is another reason other loras will need more training -- apart from variety. When movement is important, there will be many ways the items in the scene could move to validly hit the requirement of the lora, and many will need to be shown to the model to understand what is reasonable (even though it's understanding of anatomy is quite good).
Yes. Tbf its only like my 3. Lora for wan, but so far as long as you use more than 50 images it works. Sometimes motions can be jiggly though. Didnt found a reliable hoster for online video lora training. If you have a hoster id loveto test it with video
@mrkakapopoloch There are other decent lora examples in even wan that use something like 20 images and one short video, and hit pretty good results.
People use Runpod's generally, but you could rent any VM in any cloud provider and set it up with the necessary files and programs. There is definitely a lora in the last day or two where the creator links or references their runpod template.
@makiaevelio543 I am in the process of collecting clips for an experiment, where I want to see if an unusually high number of video clips (250+) can help nail down a complicated NSFW subject. Something more complex than a milking machine I mean. I have no idea if it'll work, but I do have the patience to try, at least. I guess I'll find out.
@Agkistrodon finding that balance between training data and epochs seems to be the art of the game. good luck :)
@makiaevelio543 Yep. The subject I'm going to try to train is cum-swapping (female-female). Given the huge visual variety, an unusually large dataset may be justified. I'll have to see how it goes.
是T2V还是I2V啊，480p还是720p? 能给提示词范例吗？我I2v， t2v都试了，不工作哦
Just for an example (!), here is my prompt, maybe it will help... (model:wan2.1-i2v-14b-480p-Q4_K_M.gguf)
prompt:<<Lora., The girl is on her knees, leaning forward. Her hands are behind her back, motionless. A milking machine is attached to her breast. When the device works, the girl's chest shudders rhythmically.>>
Da fuk
wan works quite well (y)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.