My first T2V test, still experimenting on that.

For generations I’m using 15–30 steps, a CFG of 7, and an 6 step shift for videos ranging from 32 to 64+ frames.

The sweet spot for weight seems to be working 0.6-1.0, got some decent results with 0.8 but I haven't tested T2V directly that much, it should be working but my experience is low quality compared to I2V.

As usual I just picked some generated images randomly from civit pushed concept loras, kudos to creators.

No trigger word used in the training. I recommend focusing on the action and little bit of general details. Milking and table are pretty common words in the training data.

Here’s an example prompt I used in one of these generations—feel free to alter it heavily:

A woman with [color] hair, [clothing and general look], is positioned under a milking table, [performing oral sex on /// grips] a large, erect, circumcised penis hanging down from above dripping cum. She is actively milking the penis as semen dripping down onto her chest. The background is white and simple.

I’m still experimenting the training process, so feel free to share your generations. This will help me gather more feedback to improve further. I didn’t have a large dataset to begin with and trained on relatively low settings if anyone’s curious.

Since this is an NSFW-capable model, please use it responsibly. No real person faces were used in the training and the rest is heavily anonymized.