What an AI dance video generator actually does
An AI dance video generator takes a still photo and a reference clip of someone dancing, extracts the pose from every frame of the clip, and generates a new video of the person in the photo holding each of those poses in turn.
It is pose transfer, not animation
The word "animate" suggests the photo is being stretched and warped. It is not. Each output frame is generated from scratch, conditioned on two things: a skeleton lifted from the reference clip, and the identity in your photo.
That is why the output can show you from a slightly different angle than the photo did, and why the model can be wrong about what the back of your jacket looks like — it has never seen the back of your jacket, so it makes one up.
What separates a good generator from a bad one
Identity hold. A weak model drifts: by second eight the face is a cousin of the face you uploaded. Hold is the single thing worth testing first, and the way to test it is a long clip, not a short one.
Hands, feet and cloth. These are the three places every model is weakest, because they move fastest and have the least surface area to be right about. A generator that keeps five fingers through a spin is doing something the cheap ones are not.
Motion fidelity. Some models smooth the reference into something generically dance-like. If the template has a sharp beat and the output does not, the model is interpolating rather than following.
What none of them do well yet
Two people in one frame. Multiple subjects confuse the pose extraction and the identities swap.
Text. A logo on a shirt or a sign in the background will be legible in frame one and nonsense by frame thirty. Choose a plain shirt if a readable logo matters to you.