I mentioned that in the model description. It doesn't work with this model. I'm not sure if it's possible without regional prompting in wan.

@clevenger Ok, I was wrong. I had an idea and it worked.
This post should be visible soon as an example: https://civitai.com/posts/26235330
You can make videos with Supergirl and Wonder Woman, but probably not Powergirl, because wan knows who those characters are and can do some of the work rendering them in costume without help. This LoRA just makes their costume and appearance more consistent. So I prompted a video of the two of them in a bar and I did NOT use the keywords for the lora. I called them by their regular names and didn't trigger the model with any specific character. The model costumes and general appearance still come through, and both women have distinct costumes. I'm pretty sure it's required that you do not use the trigger word at all, which is why I say Powergirl probably won't work since I don't think that's a character that wan knows about. The only way to get her character is with the help of a lora or at least a whole lot of careful prompting.

@clevenger Nice work! I'm impressed that the costumes and so on turned out cleanly with all that prompting.
I use lightning loras all the time and the motion seems to be ok for me, but I never just do 4 steps. I always do 4+4 steps minimum and sometimes 5+5 to help out a video that has an extra arm or something, usually an extra high and low step will fix that.

@clevenger I was able to re-create your video using some of my favorite settings, like lightning loras, samplers and so on, and the action and movement are very natural. I added slightly to your prompt and you're right that with enough costume help, you can get pretty good results. This one isn't perfect, but I didn't have the patience to try for perfect.
If you want to see the settings I used, they're basically the same stuff mentioned in my ComfyUI settings article. A few extra steps and the res_multistep sampler help tremendously, I think. Plus that high noise MOE lightning lora is the best one I've used by far, so definitely give that a try.
https://civitai.com/images/119329900

@clevenger Your videos are visible, but I think you only tagged the high or low noise version, don't remember which. So if you don't look in the right gallery, it's not there. You could edit the post and add the other model to your videos and then it'll show up in both places.

@clevenger yes, I see it there now.

Generally speaking Wan has figured it out from the training images. Sometimes you will see elements of someone else's costume, like Wonder Woman with a cape or a skirt seems to be the most common.
In that case, you can just give extra detail about the costume and it usually fixes it. If Wonder Woman has a cape, I just say "she has bare shoulders" or something similar. I gave costume/character description guidance in the model description so you can copy/paste it into your prompt if it helps.
If you look at the sample videos, I included all the prompt text for all the videos, so you can see what I had to do for each one.

@qiaoy123 Are you talking about in LoRAs you've trained, or are you saying you have that problem when using this model?

@qiaoy123 Are you describing the outfit in your training captions? Because if so, that's likely your problem. You need to make the trigger word the one magic word that includes the whole outfit, which you can only do by NOT describing the outfit in your captions. I'll put another message here in a bit with a copy of one of my training image captions so you see what I mean.

@qiaoy123 Here is the text from a Supergirl image. The other character images follow the same general format.
"supergrl. the image shows a young woman standing on a city street. she is wearing a supergrl costume. the woman is posing with her hands on her hips and is smiling at the camera. the background shows a cityscape with buildings and a traffic light pole."
In this text, I don't say anything at all about her costume or her hair or any other specific features of the person. The keyword "supergrl" is a word that Wan doesn't know, so Wan eventually learns that this word must mean "all the other things in the picture that are not described in the caption". This pattern works best when you spend a lot of effort describing EVERYTHING else in the image (or in the sample video if you use a video clip), but that's really time consuming so I usually don't go to that much effort. Wan still learns pretty well in the models I've made this way so far.
Anyway, in this setup, the word "wonderwmn" eventually is learned to mean "picture of this athletic black-haired woman with this specific costume" since her hair and costume are very consistent in all the training images. The same is true for the other characters.

@qiaoy123 Regarding your school uniform example, in this superhero dataset, I had about 20 images of each character and they are wearing the same clothes in all of those images. Then I did enough repeats to show those images to Wan for over 1000 steps. I forget the exact number for the epoch I published. So if you only show that outfit to wan once or twice, that might not be enough for it to learn very well.

@qiaoy123 You're probably right. This is why most lora keywords are weird abbreviations and mixes of letters and numbers. It's not because we're all nerds using l33t sp34k, it's because people are trying to avoid using words in the dictionary.
Good luck!

@qiaoy123 I'm not familiar with that setting, unfortunately, so I can't tell you if it would help. I've never messed with that value myself so I don't know if it'll help in your situation.
It's really annoying that learning this stuff by trial and error is so damn slow unless you pay for a really nice virtual machine on runpod or something and there are so many different values to change, any of which might make a huge difference. Lora training is kinda painful sometimes.

I hope it works well for you!