Its a penis!
Still early in lora stages, and replacing the mess that is base anatomy is taking awhile.
It appears the lora's trained on full FP16 models, don't work correctly on distilled or scaled models, seems fine running with lightx2v though.
This is more or less a beta ATM, its working decently, but does still have some janky penis from time to time.
Right now, this is only fully known to work (mostly) on the 720p FP16 model, 480p FP16 might be fine as well, didn't test on my end though.
You can try on distilled or scaled models, but don't expect good results, the blocks on distilled/scaled don't match the FP16 models.
Also noticed weird results with CFG less then 2 and steps less then 24.
Description
Trained on T2V 720p FP16
FAQ
Comments (3)
i will be running some specific for I2V and distilled models this week, right now im doing various testing on different settings to find the fastest/stablest training on a 4090
Please tell us how to easy train hyv1.5 lora?
Its not easy atm, at least not as of friday.
This is op, btw, just have different acount on my phone.
Right now, i only have slightly stable image based training, again, as of friday, may have changed over the weekend.
And its only through musubi tuner, and there training is a bit odd, not even sure if its true lora, might be lycoris converted to lora.
Biggest issue i face atm, strong back ground bleed, the background tends to train faster then what your actually trying to train, i did some runs on jenny (trained her on hyv1) and almosy every generation was in a pawn shop, even if i prompt something else. And not all images i trained on had that, but the fact that the pawn shop was present in more then 1 image, it got reinforced hard, overfit that background before getting clear training of jenny.
Again, this was as of friday, havent touched this all weekend, i know hyv released official lora training on thursday? So may have been some major changes over the weekend.