EDIT: 9/30/23 The only limit is really is the consistency over time but that we think is solvable, but with enough controlnet refinement anything is pretty much possible. you just have to use controlnets to guide towards what you want, and loosen them as you go. I definitely recommend using this despite it being trained on some nsfw content, as you can see above, its very easy to prompt negative nsfw to avoid any of it, I'm just trying to be transparent on its origins. I'll still leave it marked for intended for mature themes.

EXPANDING ON CONTROLNETS: you need them for the motion you see in my top examples, basically what the value is here is that its easy to maintain consistency in the reality around the characters, unlike base mm14, and even mm15v2

A couple different Models I have attempted to train have led me to a couple results that people seem to be interested in, though I'd still say this is an unsolved science, but there seems to be a noticeable change in the way they interact with general motion but also the style of trained data. they aren't perfect, at least I don't think so, they are still trained on incredibly small datasets

OH I SHOULD INCLUDE THIS, it doesn't have to be anime, its motion, you just choose what model you want, I just happen to use it for this model and lora, but its trained on actual porn so, its not just limited to this character.

OH AND DOUBLE ALSO: the motion is really good because I am using animatediff-cli-travel-prompt with controlnets for assistance, whats notable about this model is the stability of the image for a longer duration than normal motion model v14. but if you put in the work I'm not sure what its limit is and what it isn't capable of doing.

Maybe I'll make a tutorial for tomorrow to share. Support isn't required but appreciated :3