v1.1 UPDATE!!!

It's been a while! This was my first Wan lora a few months ago and it still gets a lot of use (thank you all for that!) so I thought it's time to make an update.

This version is again I2V and improved in all areas, longer training at a higher resolution, improved dataset and captions. You can now somewhat control where the cum lands by adding:

which lands on her {thighs|pelvis|stomach|face|breasts}.

to the end of the prompt. Other prompts are the same as the previous version. This update seems to have more realistic cum that lands better on the body or other objects.

Workflow I used has been added to the metadata of the videos.

Original Notes

LoRA for the Wan 14B I2V 480p model, untested for T2V and 720p.

Diffusion-pipe LoRA trained with 20 videos for 3k steps, 65 epochs. A few of the sample videos have my workflow attached, nothing fancy but I highly recommend the new Skip Layer Guidance for wan, either from KJNodes or WanVideoWrapper depending on if you use native or Kijai.

Samples are made from images generated in SDXL then run through wan, no upscale or interpolation.

Example prompt:

A woman holding her large erect penis. She is sitting on a couch naked. She moves her hand quickly up and down the shaft of her penis and she ejaculates white sticky liquid from her penis.