New version 1/30/25
Trained on just 113 images of glass cup/tube breast pumps only. Wouldn't recommend going above .8 strength. Gets much better results when mixed with my Edge of Reality Lora (at 0.3-0.5 strength)
Trigger:
clear_breast_pump
Context / Motion Words:
{small|long} glass breast pump one on each breast
her breasts are a {red swollen|bruised dark purple} color
glass suction tubes on her breasts pulling her breasts into the glass tube and stretching and releasing them back out
her breasts pulsing as the tubes suck and release them
A lora created for a commission from a user provided training set. Trained on 35 images, honestly surprised the motion concepts came out as well as they did considering it was a static image training set, I think captioning in the intended motion helped the model learn it despite static images.
I'd say you get about a 60% success rate from this Lora. Probably won't retrain unless the user re-commissions and provides an expanded dataset with videos.
Prompting Triggers:
Primary Trigger Word: milking_machine
Pump Types ( {1|2} indicates Impact Pack Wildcard ):
{white|metal} tube milking machine
simple milking machine,
nipple pump, nipple cup vacuum pump,
breast pump, {
blank|small|long} clear glass breast pump cups,double pump nipple and breasts,
Motion: pumps on her {nipples|breasts} stretching and releasing them inside the pump
Views: close up, front view, below view, side view, top view
Breast Styling: long nipples, tiny breasts, medium breasts, huge breasts
Concepts: tied up breasts, restrained
Description
A lora created for a commission from a user provided training set (with permission to share). Trained on 35 images, honestly surprised the motion concepts came out as well as they did considering it was a static image training set, I think captioning in the intended motion helped the model learn it despite static images.
I'd say you get about a 60% success rate from this Lora. Probably won't retrain unless the user re-commissions and provides an expanded dataset with videos.
Prompting Triggers:
Primary Trigger Word: milking_machine
Pump Types ( {1|2} indicates Impact Pack Wildcard ):
{white|metal} tube milking machine
simple milking machine,
nipple pump, nipple cup vacuum pump,
breast pump, {
blank|small|long} clear glass breast pump cups,double pump nipple and breasts,
Motion: pumps on her {nipples|breasts} stretching and releasing them inside the pump
Views: close up, front view, below view, side view, top view
Breast Styling: long nipples, tiny breasts, medium breasts, huge breasts
Concepts: tied up breasts, restrained
FAQ
Comments (6)
I'm probably missing something, but when I download a vid, I don't seem to be getting the generation metadata, like I usually did with images.
Is there another way get the generation data?
That’s because I hadn’t been packing the metadata into the videos. Finally noticed the setting for it so I’ll try to include it going forward.
@TheAIDoctor nice
That's a really interesting finding if you're able to encode some sort of motion just by captioning. Are you able to separate that from the effect of prompting for motion during inference? Either way it bodes well for training on images. I guess I'll... just have to try it out for myself and see.
I'm not sure I understand what you mean by "Are you able to separate that from the effect of prompting for motion during inference"
@TheAIDoctor well I may not know what I’m talking about either lol. But you said “I think captioning in the intended motion helped the model learn it despite static images.” If it learned motion by captioning during training that’s interesting. I wonder if the motion we’re seeing is just from describing the motion in the prompt when generating, or if it needs both. If the motion description was in the prompt but not the caption would it work as well? Maybe not, though it might be hard to know.