Perky breasts beware! This LoRA will induce dramatic sagging
Following the guidance from this thread https://www.reddit.com/r/StableDiffusion/comments/1lzilsv/stepbystep_instructions_to_train_your_own_t2v_wan/ to get Musubi tuner up and running on a 16GB VRAM setup, then this comment in particular for WAN 2.2 was essential to combating nonstop OOM errors: https://www.reddit.com/r/StableDiffusion/comments/1lzilsv/comment/nbgn4xm/
High model trained only, gets the results I'm after without needing Low. Trained against T2V model but I primarily use for I2V
Trigger word is "brstdfl8". Adding "Her breasts sag rapidly. Her breasts are now empty." afterwards can help strengthen the effect
V1 - Original release: Very excitable jiggle physics but can be unpredictable, insisting on breast expansion first and then running out of time during video gen to complete deflation effect.
V2 - Updated: More subdued effect with less breast jiggle, but seems to be more reliable at triggering the deflation effect. Added descriptor "shrivel" to training to give a more wrinkled result. Full training phrase "her breasts sag and shrivel rapidly. Her breasts are now empty"
Description
Noticed some odd behavior at times with the v1 model, where it would wait too long to start the deflate motion, or insist on expansion first with overexaggerated motion.
V2 is a lot more restrained and gives less breast bounce, but seems to be more consistent.
Have also added in a new descriptor keyword "shrivel" during training to give a more dramatic deflated look. Full training phrase "Her breasts sag and shrivel rapidly. Her breasts are now empty."
FAQ
Comments (4)
Why would You do such a thing ???
I deleted my last comment 'I can't get this to work' because I figured it out. Setting the strength of the lora to 3 dramatically improved results.
I use a low vram setup (Q4 gguf with 4 step lightning lora) which could be the issue although I use other loras at strength 1 and they work. Anyhoo it works great now I have figured that part, many thanks! Great job!
Glad you figured it out, and that setting it so high didn't fry anything! I did have issues with the lightning lora killing other motion in the videos but for me having strength at 1 still works at FP8
A great find! Working with the DaSiWa models (3 k-sampler workflow) I also didn't see expected outputs with the lora set to sane values. Using 2.0 already did great, while other loras do tend to burn when set that strong.