Recommended LoRa weight: 1 High Noise:
Shift: 7~8
CFG: 3~4 (1~2 with Lightning LoRa)
Low Noise:
Shift: 5~8
CFG: 3~4 (1~2 with Lightning LoRa)
This version uses the same datasets as the 2.1 version, with added datasets about penises.
I don't think the model is perfect yet, but it seems improved compared to the 2.1 version:
Better generalisation (in my opinion)
Better saliva supports (though not on ******** scenes, I wish to train a version for this)
Way better subject creation (adding a penis, a girl, a man to an image). You shouldn't have weird heads holding an half-penis in the mouth appearing now... At least not much.
Penises behaves now more naturally when handled (they were often too static to my taste)
Known issues:
Nude woman added in the video will very probably have a penis, reducing Low Noise weight could works sometimes, but I'll need to retrain it to differentiate penises & vajayjay (it's an edge case though, so not that important)
T2V ?
This should works also in T2V, though it's not ideal (penises are quite ugly). Maybe I'll have time to train one this year ? o_o
Low Noise ? High Noise ?
The High Noise LoRa contains most of the deepthroat movement, and basic details. The Low Noise refines the movement and improves overall details of the action (penis, testicules, penis/mouth interaction, saliva, micro-movements, etc). It means that the Low Noise LoRa is not mandatory, especially in use cases where the deepthroat is not done on a penis. So don't forget to try without it!
Lightning LoRa ?
This LoRa behaves very well with Kijai's 4-Step Lightning LoRa (CFG: 1/2, Shift: 7/8) I even consider results better with it... But that's maybe because I'm not so good at inference on raw Wan 2.2 yet.
All in all, please let me know of any weird quirks about this version, I've tested it quite a lot, but who knows. Enjoy ! (Don't use this to do illegal sh*ts, I condemn them firmly)
Trigger words (at the start of the prompt):
blowjob, deepthroat;
Helping trigger words
For POV (not mandatory but can help)
blowjob, deepthroat, pov;
For upside down deepthroat (not mandatory but can help):
blowjob, deepthroat, reverse;
For on lap deepthroat (can help, but this use case is under trained so I can't guarantee a good inference)
blowjob, deepthroat, on-lap;
For 69 (not mandatory, but can help):
blowjob, deepthroat, sixnine;
About penises, you can have moderate control on it with those keywords:
penis, dick, erect, flaccid, the foreskin is pulled back, the foreskin is pulled up, glans is visible, glans is hidden, testicules, ...
Helping phrases:
a video on a woman performing a deepthroat blowjob. [...]
A man is moving his penis back and forth in the woman mouth and throat
[...] the penis is going deep into her mouth and throat
The woman is bobbing her head back and forth while sucking the penis
[...] The woman makes the penis wet with translucent saliva [...] a bit of foamy translucent saliva accumulates on the penis
(careful with the foamy saliva, it can get weird quick)
To help trigger throat bulging (still unstable, couldn't make it works correctly yet):
Her throat is (prominently|slightly) bulging as the man's penis penetrates her mouth and throat
Example prompt:
blowjob, deepthroat; a video on a woman with fair skin performing a deepthroat blowjob. She kneeling in front of a man. The man is moving his erect penis back and forth in the woman mouth and throat, while she is sucking the man's penis, the penis is going deep into her mouth and throat. Her throat is prominently bulging as the man's penis penetrates her mouth and throat. She makes the penis wet with translucent saliva.