Indeed, what they're already doing is already hobbling the models.
Emad is right that we learn new things from the creativity unleashed by accessible models that can be run (and even fine tuned) or consumer hardware.
But judging from what people post, one thing we learn is that it seems models fine tuned on porn (such as the notorious f222 and its derivative Hassan's blend) can be quite a bit better at non-porn generation of diverse, photorealistic faces and hands too.
I'm not sure I understand this. A possible implementation could be a neural net that blanked the screen with a frown face any time it detected something it thinks was "bad". What purpose/need would pre-screening serve?
I think this is making the assumption that all frames are blocked.
> then it would be disastrous for latency
We're talking about the future here. I'm not sure it makes sense to use current tech to say it's not going to happen, or come up with latency numbers. But, "real time" inference is definitely a possibility, and is in active use for video moderation (Youtube, etc) and object detection (Tesla, etc). Nobody will notice a system running at 700fps.