User request for Phut Hon dance lora. Trying to train hunyuan video on four dance moves from Phut Hon dance videos. Dancing with arms up, hands behind head. Dancing with arms low, hands clasped in front. Dancing with arms crossed under breasts. Dancing with arms t, one arm crossed, other arm straight down. Lora was also trained with 'phut hon, phut hon dance, [arms (up, down, crossed, t)]'.
arms up - works consistently
arms down - hit/miss, add arms down, hands together, elbows straight for more consistency
arms crossed - works consistently
arms t - very hit/miss, add left arm straight down, right arm crossed grasping left elbow for more consistency
for additional visual stimulation - add Dancing with breast bouncing lora https://civarchive.com/models/1052680/dancing-with-breasts-bouncing-hunyuan-video
and/or Exotic dancer lora https://civarchive.com/models/1214079?modelVersionId=1474788
at low strengths.
NOTE: final trained version turned out better than expected.
I used this process for local training:
https://civarchive.com/articles/9798/training-a-lora-for-hunyuan-video-on-windows
vers 1.0 - 26 epochs 3900 steps.
vers .16 - 20x65 frame clips, 256x256 pixels, 16 epochs. training was pretty good learning body movement i was trying to capture.
Description
early version. trigger word 'ph2t-h0n'. other words to influence dance moves: 'arms up', 'arms down', 'arms crossed', 'arms t'
if facial features are not prompted 'blurred' face from training data may generate.
FAQ
Comments (12)
Nice lora!
Using https://github.com/ORB-HD/deface can prevent facial learning, which may help create a more useful LoRA. Additionally, diversifying the dataset can help avoid body shape fixation.
i used deface, but yes, had only one model. i was just noticing body shape when trying img2vid... thx for suggestion.
Love the concept and girl's doing the dance. However... It looks like you needed to correct for your frames per second of your clips. Everything came out in slow motion. Did you standardize all your fps between clips?
Seconded. It's just not catching that rhythm most likely due to the slow motion.
so training dataset is all 24fps. generations are 12fps interpolated to 24fps. i'll try 24fps generations. see if that helps.
i was just trying to create longer videos... :(
yeah, generating in 24fps makes a big difference...
thanks for catching that. redid samples at 24fps.
How many times did you hear the song while gathering clips? (I like it, but I can imagine hours of it)
it wasn't too bad. after first pass i mute audio. but yeah. :)
This dance is not from that meme, is from the mememe animation
all i know is a user sent me a youtube video of a girl doing these 4 dance moves and wanted me to make a lora for it.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.