If you run into issues with the output being affected, please try with/without your attention-modifying nodes.

I use res_multistep sampler and simple scheduler.

Tons of workflows add way too many attention nodes that modify how the base model works, which can invalidate many of the training weights, especially on pruned models.

The most recent model has been generated for MiniMax h3. Audio was not trained with this dataset, at all. Penises were also not trained very well. This allows you to give any cylindrical object a titjob, but may require other LoRAs or a reference image for penises.

This will work with or without Turbo LoRA for both ref2video and txt2video.

Using a phrase that includes the word "titjob" or "titfuck" works best.

Future plans: It tends to collapse towards a select few angles, male parts are often misshapen, and the model influences breast size. I plan to improve the angles and breast size. I'll work on a penis LoRA that plays nicely with this one in the future.