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
interim test version. only 5 epochs. model faces blurred during training to see if face loras work better. v2.0 will be uploaded when reaches 10 epochs.
FAQ
Comments (14)
Comment says this one is 5 epochs, next is 10, how long do these take to train?
i'm currently running on an amd 7900x w/128GB dram and rtx 4090 w/24GB vram. about 20-24hrs runtime. i checkpoint every 2hrs, so it takes me a couple of days...
that's with 37x2sec video clips sampled at 460x460. it'll be about 3400 passes when finished.
@tedbiv So I assume you go around finding clips of the action you like and cropping the videos in to show just that, then scale down to 460x460? Any chance you could post dataset and training data for us mortals as an example?
i have a 6 minute video of this woman dancing. https://civitai.com/images/50237696 i can't post the video, it's copyrighted. i use this process https://civitai.com/articles/9798/training-a-lora-for-hunyuan-video-on-windows and copied dataset creation from this guy. there's a couple of comfui workflows on creating the 2 sec video clips and text tagging them. i create dataset at 1024x1024, then scale it down when training in the scripts.
@cexoga8654284dfdfd but yes, i find video clips of the motion i want, just recently tried blurring the face, cut it into 60frame 2 second clips, create the text tag files and run the training scripts. i'm fairly new to this these are my first 2 loras i've tried...
@tedbiv Well you're going a great job. Thanks, I'll check the links.
thanks, it's a learning process. but i like a good exotic dancer... :)
Epoch doesn't mean anything, because the depends on number of repeats, grad_accum and other factors.
How many total steps are you at? what is your dataset size?
@azeli 37x2sec clips. This version is around 1644 steps
@tedbiv how you managed fit 460x460 in 24gb vram.
How bit swap you set in toml and accumulation?
@CyberAImania i started at 512x512 it overflowed into dram. so i watched taskmgr gpu vram to see how low i needed to go. 460x460 fit but it was slow as hell. now i run between 256-320 depending on how many clips i have.
@tedbiv Thanks for the reply. That’s what I thought too, since I tried fitting 480 x 480 myself — it worked, but only with:
blocks_to_swap = 30
gradient_accumulation_steps = 8
and that makes the training time too long for me... a waste of time and electricity. It's better to rent MassedCompute or RunPod if you really need higher resolution. Locally, on a 4090, I think 256px is the sweet spot.
@CyberAImania agree. i am currently only able to use 170x170 for wan training... debugging why.