Almost forgot.  Minimax H3 is surprisingly good at the entranced voice.  The problem is pacing.  If you let Minimax handle everything, you'll get those pregnant pauses AI is famous for.
I want to edit the audio separately, so the pacing is good.  Then offer the audio as reference. 

If Minimax really is best for this, I'll simply render scenes with minimal video, then use that audio.

Nice work!  Blank stares are my thing and you delivered!
Adding them being lit by whatever is hypnotizing them is a nice touch.

Has anyone experimented with those little art manikins for poses and such? I know that it isn't that hard to find reference photos for poses, but I think that the fine tuned control could be useful when you have a very particular vision in mind.

I have not! But it is an interesting idea. I know there are also these "anatomy reference" websites for visual artists, which are basically just 10s of thousands of renderings of various body types in every pose imaginable, from every angle imaginable. So that might be useful here. Another possibility I've entertained (because I happen to have one of those digital drawing tablets) would be to do quick little stick doodles showing the exact poses/scene layout I want, along with a generic character reference, and perhaps a generic reference for the look of the room, and feed it all into the AI with instructions similar to yours. This would be particularly useful if a person actually did have a "directorial vision" in mind, rather than just letting the AI come up with whatever kind of shot it wanted. And perhaps it would help with pesky little details that AIs seem to consistently mess up even with plenty of hand holding (for instance, getting a character to look at the exact right point on a hypnotic object - apparently this is hard for them to get 100% right!)

The other thing I wonder here, is that perhaps these methods would be a sneaky way to get around certain content filters. For instance with your example, just asking for a picture of an anime girl groping her own boobs might throw some AIs into a fit, but showing a posed manikin & saying "just make it like this" might get through fine

Perchance is really nice for playing around with a low-end AI without any kind of daily limits.

On the image generators, in my experience I've found that it does closeups and half-body shots reasonably well, but if you try to do a full body pic it tends to have a hard time with the eyes. I've also only had limited success getting whitewashed eyes on realistic humans, though changing the settings to some form of animated style generally works better.

Of course, it is also possible that I just suck at creating prompts.

https://perchance.org/ai-chat and https://perchance.org/ai-character-chat are story/chat generators. The first one is decent for simple conversation-type interactions. You can add additional characters but it doesn't really like it. The second one is made for multiple character interactions as well as simple conversations, though even with the "long-term memory" function enabled, it really starts to lose the plot once it hits its token limit (whatever that is, I haven't tried to calculate it.)