H3:

Please understand that this is a highly experimental early Lora with a lot of issues, but since it took me almost the whole day to train, I decided to upload it anyways. You know - for science...

This lora was trained as I2V but seems to work in the R2V reference model as well. I have not tested T2V, but would expect some messed up potato-penises.

There are several obvious problems: Way too runny semen, some completely ridiculous liquid bubbly sounds from the semen and stroking, and I generally feel the quality deteriorates when using this Lora. (or any H3 Lora, at the moment)

But this is what we have until our unsung genius heroes figure out how to fix the kinks with H3 Lora training.

Please check the prompts on my examples for inspiration.

I used this Lora at strength 1.0 on all examples.
(Lower than strength 1.0 will roll it back to H3's default paint bucket cumshot)

LTX V1 vs Beta versions:
I 'm honestly not quite sure what the differences are between the 2 versions, but V1 was trained with double the data, so I suggest you try both and see what fits your needs.

This is not a standalone Lora, but meant to be used with other Loras or checkpoints, like the amazing LTX2.3 10Eros checkpoint.

Prompting idea, even though I am sure it can be prompted better:

CUMSH0T.

The woman is stroking the mans penis up and down with one hand.

Continuous short pulses of long ropes of white semen shoots from the tip of the penis which hits her face, while she is moaning.

The semen hits her forehead, nose, cheeks and mouth, the semen is running down her face and body.

The penis is shooting long ropes of white semen across her face, nose and mouth, with long streaks running down her chin and tongue throughout the cumshot.

This Lora can also be used if there is no penis in the first frame - prompt it somehing like this:

A large erect penis enters from the left side of the frame.

The man is stroking his penis up and down with one hand.

V1 Training: 3000 steps with 40 (768x768) clips at 5 seconds / 16 fps.
Beta Training: 3000 steps with 20 (768x768) clips at 5 seconds / 16 fps.

Even though this is trained as an I2V model, it seems to have effect on T2V as well.

Known problems:
- I see a colorshift in the first seconds, and not sure yet, if it is because of overtraining or my workflow.
- Cum is magically appearing all over the body very fast, but hope it can be fixed with more training data and better captioning.