Minimax H3:

H3_0.1: Strength 1. Generally recommend to mention "an erect penis" for anatomy.

Not as good as the wan builds, but better than feared. ref2v also works (but worse).

Different prompts than WAN, some other pointers:

thick semen shoots straight upward in one single jet, landing deep in her hair

for consistency sometimes some attempts like milky-white, slightly translucent semen shoots in an arc, landing on her nose

semen shoots across her nose and cheeks in thin streaks, a thin drip running down

Also should take direction, e.g. thick semen shoots straight upward into her hair twice" / across her eyes / thick semen first lands on her eye, then a jet curves back into her hair, a third jet shoots straight into her hair, and a final jet lands again on her eye

WAN2.2

For I2V probably 0.3 is best.
for T2V/else I think I still mostly like v0.1 over 0.2, both probably have their own things they do best.

v0.3 (I2V):

Slightly reduced dataset ("spray" not contained anymore wont work!),
Lower rank, higher resolution, different training.

I2V only. high/low roughly 1.0-1.5/0.3-1,
good start likely is 1.3/0.7 , 1.35/0.72 , 1.4/0.75 / 1.5/0.6 etc.

v0.2:

Slightly optimized dataset, rank 32 (lower filesize).

Stock Wan2.2 T2V strength 1/1 or 1.5-2/0.4-0.9 (e.g 1.6/0.42) as a starting point. Works for I2V. Try out.

v0.1:

Stock Wan2.2 T2V LoRA strength roughly for high/low 1.2/0.55 as a start point.

Works for I2V at maybe more like 1.4/0.6
(for quick test used samplers 0 to 4 and 4 to 20 steps wise).
Let me know if you find better settings...

Recommending more or equal to 8 steps, e.g. steps=20.

For more prompts select the low version and download the training data/images, which contains a text file with all prompts used for the dataset.

Example prompt:

cuminhair An erect penis gives her a cumshot and shoots cum into her hair.

cuminhair An erect penis gives her a cumshot and sprays cum into her hair.