Yeah, it's honestly annoying how pervasive porn models are. I think like 90% of people using SD are using it to make porn or waifus.
As I was driving home last night... you know how when Szilard was walking home after hearing Rutherford be dismissive of nuclear chain reactions, stopped at a stop light, and suddenly the idea of neutron chain reactions hit him like a ton of bricks and it was like the world peeled away around him? I had one of those moments when I realized that all of the components are now out there:
Story Diffusion (e.g. "Open Sora")** Generate realistic videos of anything, from input commands.
MotionCtrl (ability to direct objects and camera positions in AI vids)
Open LLMs (countless), trained to roleplay multiple characters, control scenes, and to issue external commands
VR headsets (with a mic + voice recognition and/or hand controls; or alternatively, LMMs can accept voice natively)
Optional: FPV cameras on the headset, plus depth map model (Midas, etc) background stripping, plus vid2vid to integrate the user's body into the scene
And... remote-control sex toys that accept external commands.
You can see what someone is surely going to put together from these pieces over subsequent years: a porn holodeck. Where an open LLM or LMM takes user inputs (such as speech or motion) once every 1-2 seconds, determines character reactions, directs the next 1-2 seconds of video generation, generates audio, and when appropriate, issues commands to control... peripherals. It'll take quite a bit of compute power to handle the video generation all in realtime, but you can fan it out to multiple cards, so it's doable if you keep the resolution and steps down & just AI upscale the outputs and motion-interpolate between frames. The LLM/LMM wouldn't need a huge number of parameters for such a role - even something like LLaMA 8B as a base model should suffice well; there's no issues with running that in realtime on any semi-modern GPU.
I'm like 80% ACE so the whole thing sounds rather gross to me, but knowing how people are using SD already, you know this is... sigh, I walked right into the pun, didn't I?.... "you know this is coming." :P I guess it's good that the perpetually horny will have such an outlet.
** - Story Diffusion is "out". but they haven't released the video-generation half of the model yet, only the keyframe-generation "comic book mode", but they reportedly plan to release the whole thing; there's also less advanced models like Stable Video Diffusion and the like (which MotionCtrl was designed for).