To the mod team/TOS disclaimer

This model complies with the TOS to the best of my knowledge. For it to operate the person must already be in a sexual scenario(eg a woman and testicles must be in the scene) or other loras must be used to move the person into the scenario (such is the case in my examples). It, by itself, does not add a penis/testicles into the scene. None of the training material included anything that took a person and moved them into a sexual scenario, it all started with both man and woman in the scene (like adam and eve ;). Happy to show the training material to the mod team if that helps at all.

Any models that i have that don't comply with the Civitai TOS are hosted over at hugging face. The link is in my profile, and might be worth a look ;)

Concept

This is an updated version of my POV ball busting model but with some things added and one thing removed. I've added 3rd person perspective ability but removed the ability to slap (how can she slap!?). I found having both action concepts was dilluting the concept of punching. Maybe that'll be a seperate model at some point.

Usage

Base prompt is:

A video of a woman punching a man in the testicles.

By default this should just work with most angles, but for POV add this:

The camera angle is from the man's POV.

I find sometimes adding:

The woman makes a fist with her hands.

Can help her get a better gen. This model should work fairly well with many other anatomy loras, as per the examples.

Weight of 1.0 for both high and low should work well. Sampler wise most should work but i'd avoid unipc as it does some weird things. Stick with Euler/LCM/Res.

I highly recommend using @thomas852002887 flaccid penis lora in combination with my soft suck lora. The combo gives great flaccid pp physics. Equally @bial rim job lora can really help get the shape/movement of balls from behind.

Training

The original version of this lora was my first every wan lora, and i've learned a few things since then! So while v1 was trained both high and low at 0.4 MP with 6 frames. v2 has been trained at 0.06 MP @ 6 frames for high and 0.4 MP @ 2 frames for low. Batch of 4 on high and 2 on low. Overall I have found the punches can be a little slower,but acceptable. Low might need to be trained with more frames but will need more testing. I'd be curious to see other peoples findings. Dataset used was 34 POV videos and 22 non-POV.

Captioning was automated using JoyCaption and DarkAges Llama 70B. Trained on diffusion pipe on 3x3090s and 1x5060ti.