So my process is about the same for all of my loras, so as far as why this one would be better or different from others, I have no idea. When I rip a video from a website, i choose the 480p option for the download. Then I crop it to 3 seconds, making sure the movement is clearly visible in the first 2 seconds of the video. Then I crop it to 640x480 or 480x480 if needed to get rid of any watermarks. Otherwise, I don't need to crop it. Then I convert it to 24fps, just to make sure that is the frame rate of the video. Then I just manually caption it.

At that point, I specify 244 as the resolution of the video in the config file, and set the frame buckets to [1, 50]. (You don't need high resolution video to train a movement, it seems). If you download the training data from my missionary lora, the settings should be almost exactly like what I did here. I may have slightly messed with the learning rate, but I honestly haven't figured out how much of an impact that has on anything.

I suspect what ultimately determines how good the lora ends up being has a lot to do with how much the base model already understands about what you're trying to do. I think one difference between my training data for this one and others is I had very few with faces in view for this one. I don't know if that's a good or a bad thing.

I actually put together another version of this, where I added a whole other directory of just images of a wide variety of women, captioned without the trigger word just to see if that would help improve facial variety and make character loras work better. I never uploaded it because I honestly don't think it made any difference at all.

My dataset wasn't very large, but it was definitely multiple people. It's funny, you're the second person that has mentioned faces as a problem, but I haven't experienced this myself. I even tried a couple methods to improve the faces, but I haven't uploaded those versions because I see no noticeable improvement. Mostly because v1.0 is already so good to start out with, as far as I can tell.

The movement still works well if you use a weight of 0.77, so I would recommend trying that and seeing if it helps with what you're trying to do.

@spike4379 Yeah, I wish I knew. The last attempt to train this literally had no faces at all in the dataset, and there was no change, so not really sure what to do about it. You could try adding countries of origin and random first names and last names to see if that helps anything.

My missionary lora definitely was affecting faces negatively and not respecting character loras very well and training without faces had a very obvious improvement. With this one, character loras are already working well for me and there's no improvement. I wonder if there's some difference in workflow that is making it not work well for some people.

Yeah, I'm in the process of creating these loras for Wan as well. It'll take a little while to get through them.