This lora attempts to capture 'handsy' caressing action of exotic dancers. Caressing their body, legs, breasts hips and more. Trigger word 'br38st-m8ss8g3'. Also trained with 'rubbing breasts, caressing breasts, massage breast, fondle breast'. Can be used with https://civarchive.com/models/1214079/exotic-dancer-yet-another-sexy-dancer-hunyuan-lora
to enhance the overall 'Exotic Dancer' experience.
NOTE: if used with 'Exotic Dancer', loras strengths will need to be lowered ~0.3-0.5 to have any physical characteristic from prompt used due to same model used for training.
Wan2.1 version trained with Wan2.1-T2V-14B, also works with I2V model. Might need to add 'massaging breasts' to motivate the lora...
Description
initial version
FAQ
Comments (23)
I think you've trained in a face, as I keep getting the same face over and over, about 50-75% of the time.
edit: I see the face in the samples here as well! Shame as liked the lora but it's useless if it doesn't allow other faces in.
thx for the feedback, i noticed that also. i'm new to loras. how should i retrain the lora and not use the face? should i just crop faces out of video clips? so only body and motions are visible?
plus, the video clips are all from same model.
ok, i found a way to blur the face, i can re-run the training. i'll see if that helps.
@tedbiv awesome, yes I personally haven't got to training motion (video) yet, but I did see someone else here post something about a script to blur faces, assume that's what you've found
looking forward to v2!
@azeli actually one of the tools i use 'davinci resolve' has a tool to select an area to blur then tracks that area in the video. i'm about half way through training. we'll see if it's any better. maybe tomorrow.
just to add to the conversation, my understanding is that you want multiple models of various shapes and sizes for the lora to learn the concept rather just that 'person'
@az420 100% for this type of lora yes, dataset should be ideally all different women unless you want the breasts to be exactly the same as well
@azeli i agree, but it's difficult to find good video content. i was mainly using this as a learning tool to build loras. most sites won't let you download content anymore.
@tedbiv Think you've just hidden the fact the face is trained in by using the disney lora, i'm still getting the same issue as v1 and 1.5
@tedbiv also you should name the lora filenames as well just to make it easier to know what's what
@azeli good idea.
@azeli actually in v1.0 the model face bled through a face lora. other than trying to acquire more video, i'm not sure what to do. i only have the one model. i can try masking her face instead of blur.
@tedbiv masking does work to some extend because it technically adds to the pool of more diverse things the model hasnt seen, but other than that no, there are no shortcuts here than to simply put in the effort needed to get more high quality training data. And please do not start blurring the faces, because then you're effectively training the lora to blur faces. Remember that its more important to have a diverse set of training data than 100 videos of the same girl doing the same movements standing still in the same place.
@squawgee got it. i hadn't thought of training the lora to blur faces as a side affect.... it's hard finding good training data... :(. message received. thanks for the feedback.
@squawgee just fyi, found this nugget ... I used https://github.com/ORB-HD/deface to blur the faces in the training videos, and then added "woman with a censored and blurred face" to the captions.
@tedbiv Yeah I've trained loras with censored faces and it's fine as long as you caption it as censored/blurred face. That's a big part of captioning, to caption things that you don't want the Lora to inherently learn.
@definitelynotadog yes, i had to save 'woman with a burred and censored face' in the captioning workflow so i don't forget it... still learning about what to include/exclude in captioning files.
@tedbiv What captioning workflow do you use?
@az420 https://civitai.com/models/1129181/hunyuan-video-wan-lora-training-toolkit sometimes it generates too much... but it takes 3 images from a clip to create caption content.
@az420 there was a bug in the python script used to remove duplicate strings, let me know if it's still there i'll give you the fix if you're interested.
@tedbiv muchas gracias!
@az420 de nada