Yeah. As much as I appreciate them, I feel like there are way too many people doing straight up porn loras. Loras with more subtle seduction scenarios are a missing niche.

Btw, feel free to share more ideas like this.

I've been waiting for someone to do this since Hunyuan! Haha. I thought maybe people weren't that interested in this concept. I'm glad to know you like it.

Let me know if you have more ideas.

@makiaeveli Yeah, pretty much. The lora is inactive in the first loop.

Lol. You should still give it a go and post it. It'll generalize better mixing the loras

Me too! Wasn't expecting other people to be into it, since I've been waiting for so long for someone to do it.

Feel free to make some posts of your creations. Curious as to the prompts people come up with

Yes. This is because of an issue I had in training. I was initially training with data on both sides. The model has trouble learning that the opposite cheek needs to bulge. Sometimes both cheeks bulge. So I had to compromise and flip videos so that only the right side was ever bulging in all of them.

Let me know if there are other gestures like this and I will train

Yes. I replied to another comment with a similar observation. Will post here too for visibility. "This is because of an issue I had in training. I was initially training with data on both sides. The model has trouble learning that the opposite cheek needs to bulge. Sometimes both cheeks bulge. So I had to compromise and flip videos so that only the right side was ever bulging in all of them."

Yet, although diminished, we still have this problem you mentioned. There are certain images that the model will struggle getting the angles right. Sometimes prompting and describing which hand you want to use helps. Otherwise cropping out problematic parts of the image helps. Sometimes its just too much work and using another image is easier

@fctq9169 Yes, maybe? Its tricky because it's the direction of the motion of the hand more than which hand it is. Example, in the main example with the woman in the auditorium, she is using the right hand, but her her right cheek is bulging, not the left. The motion of the hand makes it correct, because her hand moves from her left to right.

All this model was able to learn is, one hand moves and one cheek bulges, doesn't care which one. Gets the right combination maybe 70% of the time.

Haven't tried. Doubt it. But maybe can try the low noise lora with it?

Btw, you can use the same workflow you're using with 2.1 and plug in the 2.2 WAN low noise model with this low noise lora. I think that should be the best bet if you're having problems with 2.2 VRAM requirements

Hmm. Interesting. I did not pay attention to this aspect when collecting data. The model tries to preserve the original posture of the person. Maybe start with a picture where she is showing her palm?

From your description of the output, I know what you mean. I've had that before. I get that when the lora is not properly loaded either due to the lora not being fully downloaded or the path not set. You can test this by setting a fixed seed and bypassing the lora. If you get the same video then you will need to make sure you the path to the lora is correct.

If you're using Rgthree's power lora loader, as I am in the workflow, it fails silently when even the lora not detected.

I have not experienced that. Can you describe what you're doing more? Is it I2V? What resolution are you using.

@clodius02218 All of these generations had accelerators on them. Maybe try the workflow embedded in them?

Sorry just seeing this message.

Here's a few pointers for that:
- I2V models are less prone to copying the facial characteristics of the dataset. That being said, anything not visible in the first frame but visible later in the training video (eyes closed in the image proceed to open in the video, or mouth closed in the image shows teeth later) will be learnt by the model. Think of it like anything not in the initial image, it will need to imagine to generate on its own during inference. So to do this it will need to remember what it looks like, so it ends up copying the aspects of the training subject's face.
- Try creating variations of your dataset using VACE. You just mask out irrelevant parts depending on your video like the eyes, hair etc. so that there's more variability
- Combining the two ideas above, if the first image has their mouth closed and you don't want the model to learn the teeth of the subject, you extend the video using VACE and start with a smiling picture. This way the starting picture already has the nature of the teeth in it, and it won't need to learn it

I did not use the 2nd and 3rd technique on this particular lora, but has worked for me on others. This lora only has the first aspect in play.