I'm mostly happy with this one. In typical hunyuan fashion, it doesn't always listen to directions, but most of the time you end up with a decent video. The previews were generated at 0.9 strength.
The main part of the prompt is:
A woman is straddling a man and having sex with him. She bounces up and down on top of the manDepending on the angle of the camera, you can either directly see the genitals or not. You can add this to your prompt to try to get them to show up:
You can see the woman bouncing up and down on the man's erect penis.You can prompt for different angles:
{The view is from the front and at a 45 degree angle.|The view is from the side|The view is POV}Some of the training data had the woman facing away from the camera, which you can try, but the 2 people are positioned weirdly more often than not when doing so.
By default, the woman's knees will probably be pointed down, so if you want more of a squatting position with her knees up, you can prompt for that this way:
She is squatting with her knees raisedDescription
FAQ
Comments (26)
Very nice - thanks!
One on the very best Hunyuan Loras, again! With some Loras you need to filter for double blocks, otherwise they get grey and blurry. But yours work without that somehow. Do you train on double blocks only somehow ?
Not that I know of, unless I'm somehow doing it by accident. I'm not sure how I would even do that.
So then you are just a natural talent xD ! Do you use diffusion pipe or musubi trainer? Maybe that's the difference. Just read there was a problem with the earlier Januar musubi version. Maybe that's it.... anyways, can't wait for you next Lora =)
I use diffusion pipe. I generally don't need to filter any blocks if I'm only using one lora. If I'm using a character lora + motion lora, I filter out single blocks on the character lora, and leave both block types on the motion lora.
@dtwr434 hm I even get the problem with a single Lora. But I see those slight checker board patterns that are described as part of the problem. I will retry training with updated scripts and see if it's fixed. There seems to be a big bunch of motion date in the single blocks. So that's why character Loras don't suffer from filtering, but motion Loras do =)
Excellent.
works well
Any chance of releasing the training data/dataset?
Your loras are great, any chance you'll bring these over to WAN?
I'd be willing to try, to figure out if WAN is worth it, but so far I don't know of any way of training the 14B model on my 24GB card. So, unless a method of doing that becomes available, probably not.
I've only done a little testing but this already heads and shoulders above the previously created cowgirl lora on civit. Importantly, it allows them to lean forward instead of just pogo sticking up and down with their back straight up in the air like the previous one.
this is basically perfect. Flexible, most angles covered, motion is fantastic, no specific "face" and follows prompts. Please make every sex position in your style, or at least let us know your secret!
mate this is an amazing lora as it doesnt impact other lora models, there is only one issue. whatever you try it produces very small penises. If you ever retrain if you could fix that it would be great.
I noticed that too... I guess it's the motion in the ocean... ;)
This is just absolute perfection! Your best lora ever! (Edit: The best lora ever!)
Please do your examples that retain the workflow. It's polite and helpful!
Sorry, but as long as it embeds file paths from my personal computer into the metadata, I don't really feel comfortable sharing that. I figure at this point everyone has their preferred workflow anyway, and I provided all the prompts for my images as well as advice on how to prompt it, so I would hope that's enough for people.
A. Linux because windows is terrible at that, not just here. Also good practice to not make your windows username your real name or identifiable at all. Again, not just here.
B. I am not 100% but I'm pretty sure only the relative paths are included in workflows. So if you put something in to C:/users/jimmy/comfy/loras/MyCoolLoras/lorafilename.x then only MyCoolLoras/lorafilename.x would be in the workflow. The only paths that would be included would be relative to comfy's own path.
You do you, but if you are security minded, it's time to consider both above.
@yekal394674355552 I'm on linux. It embeds the full path, and while it doesn't contain my name or anything like that, I still don't want to share it out. My workflow is super bare bones, and there are much better ones you can find on civit. I don't think it'd be of general interest to anyone, since it's pretty much the default workflow from kijai's examples. I have included the json workflow in the "Training Images" download for some of my loras, so you can go find one of those if you really want.
Great job! Your loras are awesome. I love the movement of the breasts on this one. Next request doggy from the side/front!
for WAN please..
@dtwr434 super
This is awesome work! How did you manage to build the dataset? Did you exclude the face of the person? Or blurred it?
works amazingly with framepack