I made this Lora to get more fluid movements in dances, it also helps slightly with other types of actions, it works with terms like: sexy dance, dance, belly dance, and similar
The training was with 34 videos in 24fps 55 frames, 240 res
15 Epoch, around 3000 steps, 32 rank, 5e-5, 30 repetitions
It took about 10 hours
Description
FAQ
Comments (15)
This is a good one, thanks for sharing. Since you trained it on low res videos, maybe if you top it off with a training run on 720p images it will look more clear and crisp? I think it might, based on results in other loras: https://civitai.com/models/1109266/egirlinfuencer-hyvid-hunyuan-video-lora?modelVersionId=1246337
I mean normally people would upscale any lowres sources anyways. I think the problem is about different angles, these things end up trying to find the average. You're right if you give it perfect highres sources like from a streamer will give better results, but those are also character loras -- which are far more straight forward to teach. It's mostly the face, but it'll obviously pick up the other details. Idk you prolly know all this Im just saying that I dont think it's as simple as having good sources
In my tests I have noticed that the action Loras do not affect the final result, if you generate videos in 1280x720 they will look very good, I do it at a lower resolution just for speed.
@DavidGalardi yeah, actually, now thinking about it the model should will not get visually worse with lower res videos here. The video is still crisp if not more crisp when downscaled. Maybe it even generalizes actions like this better then when given high res data to latch on to.
By the way, 10 hours on what GPU?
3000 steps, I mean that's probably consumer hardware. Decently sized dataset though, so maybe not.
Sorry, I didn't specify, I'm using a 4090, I'd like to increase the resolution but I only have one PC and I use it for work too, so I can't do both at the same time, I've tried online services but I personally prefer to do it locally
@DavidGalardi I agree completely on the principal alone, this technology feels useless if it's locked behind some supercomputer
how you do training for hunyuan can you share please!
@deimos88 I tried musubi tuner but results are scuffed, waiting for a fix. If I'd have a beefy GPU I'd try diffusion pipe.
@deimos88 With this guide: https://civitai.com/articles/9798/training-a-lora-for-hunyuan-video-on-windows
Your results are amazing.
Is the learning image only the waist? Or is it the whole body?
What was the text you described in the dataset text file to achieve this? As i'm attempting to make a dance "move" video lora for the first time, and having struggle to get the move itself. Visually, i get the character it was trained on, not so much the dance itself. Using 51 frames being 608x1080
really good. The movement is super fluid and stylish!
can you make a 'magic bomb' dance lora? for wan 2.2
if you dont know what it is search google:
magic bomb dance challenge
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.