Why AI video gets hands and faces wrong
Hands come out wrong because they are small, move fastest and can fold into a huge number of shapes with very few pixels to get right, and faces drift because nothing in the generation forces the last frame to match the first.
The hand problem
A hand occupies a tiny fraction of the frame, has more independent joints than the rest of the arm combined, and moves faster than any other part of the body during a dance. Every one of those makes it harder.
It gets worse when the hand was hidden in the source photo. If your hands were in your pockets, the model has to invent a hand and then animate the invention.
The identity drift problem
Each frame is generated conditioned on the frames near it. Small errors accumulate, and over a long enough clip the face at the end is a slightly different person from the face at the start.
Better models drift less rather than not at all. Drift is also the reason a long clip is the honest test of a model and a two-second sample is not.
What actually helps
Show the hands in the photo, away from the body, fingers relaxed and separated.
Prefer shorter clips when fidelity matters. Drift is a function of length.
On a full-body routine, try a model known for body movement before assuming the photo is the problem — and try a different photo before assuming it is not.