[rsynnott]

I'm not convinced that AI image generation _is_ getting better at this point. If anything, it seems to be getting somewhat weirder-looking.

Like, I gather that prompt adherence has improved somewhat, but the actual output still looks _very_ off.

[captainkrtek]

I think still images have improved remarkably, though some things still look off for people (they look too-flawless). Whereas video suffers from the strange fluidity and unnatural motion of things.

[rsynnott]

> though some things still look off for people (they look too-flawless)

I think _maybe_ there's an uncanny valley problem (and it may vary person to person). I found Stable Diffusion 1.5's output quite _bad_, say, but not as, I dunno, objectionable and wrong-looking as current models.

Video has always been a complete mess, remains a complete mess, I don't see any real path towards it not being a complete mess. It is, in fairness, a _much_ harder problem.

[captainkrtek]

Agreed.

I've also found the proliferation of "AI film director" created "films" funny.

The output is a strung together set of 2-3 second clips telling some story. The characters changing between clips, the scenery changing drastically, etc. There is nothing cohesive. I imagine its a similar "context window"-like problem, if you have to keep many minutes of visual context.