What is realistic today

AI video sits roughly where AI images sat two years ago. Genuinely useful, clearly limited, and badly oversold by the clips that go viral.

What works well now:

Short shots. Five to ten seconds. Beyond that most tools lose track of what things are, and a face or a logo starts drifting.

Camera moves over a still. Give it your own photograph and ask for a slow push in or a gentle pan. This is the most reliable trick in the whole category and it looks professional immediately.

B roll. Traffic, weather, an office corridor, an abstract texture behind a title. Material nobody studies closely.

Avatars reading a script. A presenter who does not exist, speaking your words, in your choice of language. Quality is now good enough for internal training and simple explainers.

What does not work yet:

Continuity. The same character across several shots, with the same face and the same shirt. Tools are adding this and it is not solved.

Hands, crowds and text on screen. Same weaknesses as still images, made worse by movement.

Anything with precise timing. Lip sync to existing audio, a product doing a specific action at a specific second.

So the honest workflow is a mixed one. Generate the shots that suit generation, film or licence the rest, and assemble everything in an ordinary editor. A video that is entirely generated usually looks it.