Made with 78 videos of lesbian cunnilingus- both faces always visible in standing sitting and laying down positions. Made for T2V, works pretty well for I2V
v0.9 and v1.0 are just different checkpoints, 1.0 may be over trained but I feel it is easier to lower strength/cfg if necessary. 0.9 is the most flexible, I like 1.0 the best
The "For Lightning" models are just overtrained versions that seem to be more compatible with speedup loras than the other versions. I have noticed it is helpful to enable cfg to ~1.3 on the high model and put "Kissing" and "She licks her own hand" in the negative prompt. I know it slows down generation but its just for the high lora and it gives better results.
examples should have workflow
[512x768] and [768x512] @ 97 frames
You can choose who is receiving/giving and what they look like with "left" and "right"
I did manage to label orgasms with "the woman on the left is having an orgasm" and differentiate "is licking" vs "goes down to lick" and label "innie pussy" in the training this time, it seems have worked
I would give more steps to the LOW model if the tongue isn't moving correctly, I would give more steps to HIGH if you want more orgasmic movement
It's not performing super well with all steps utilizing speed up loras
I like to use a mixture of Wan-Lightning lora in the beginning and end, and without speed up loras in the middle to enable cfg of about 4.0 .
examples should have workflow
Description
this version is just an overtrained version that works better with speedup loras
FAQ
Comments (3)
it has come to my attention that wan 2.2 is supposed to be at 24 fps, and I did in fact train this at 16fps
No, you heard wrong. Only the small model nobody uses was trained at 24fps or was ever intended to be 24fps. That's also the only "true" Wan2.2 whereas the 14B model these are used in is and was always a 16fps model and there's never been any confusion about that except with users who can't be bothered to RTFM and who just run other people's workflows without knowing anything about how any of this works.
Did you ever wonder why this model still uses the Wan2.1 VAE and Wan2.1 lora have always been compatible with it? Because it's an update to Wan2.1 not an actual new model. And Wan2.1 was 16fps.
Now if we could just do something about the dumbasses making "50fps" and "60fps" and "24fps" workflows for the 14b model when that's actually impossible to really do...
anyone tried it with I2V?