Baked another half-assed slop LoRA. Trained sloppily on 11 of my old AI generated videos that were using various WAN 2.1 twerk LoRAs using PGC's modified GUI of Kvento's Musubi Tuner GUI with a minor fix for WAN 2.2 I2V training that just made the --i2v toggle only apply to latent cache creation instead of training. Trained for 20 epochs but the uploaded epoch is a 50-50 merge between epoch 7 and epoch 3. Videos were lazily manually tagged as usual like A black haired woman facing away from the viewer and wearing a short white skirt with no panties twerks her ass causing her butt to shake, jiggle, and ripple. or A red haired woman wearing a hat and jean shorts twerks causing her butt cheeks to clap together. There is no relevant genital data in the dataset so use a genital LoRA if naughty bits are out of view in your initial image. Prompting something about twerking and ass shaking will work probably. There is some breast jiggle baked in despite the dataset mostly consisting of back views and side views. If the initial image is from the front, the subject may try to turn around but there was limited testing on my part or in other words, I basically only tested it like once and it's that example image of Lancer Artoria in the swimsuit that almost turns around. A lot of the test image prompts were something simple like a description of the character and then

twerks causing her ass to jiggle and shake

Didn't do much testing but if combined with Slop Bounce, probably expect giga janky slop motion. Training data includes the modified musubi_tuner_gui.py for PGC's GUI with the fixed toggle for properly caching latents for i2v as well as the .json and .toml files that contain the dataset settings and training settings. Also has the workflow images for the model page's initially uploaded videos although the videos should also have the Comfyui workflow included for once.

Slop Twerk WAN 2.2 I2V - high