One clean frame of the product on a transparent background: correct shape, correct label, correct proportions, nothing else in the picture. Every subsequent angle, set and motion clip references that file instead of a fresh prompt.
The plate is also the arbiter. When a frame is questioned, it is checked against the plate at forty per cent opacity: not against the previous frame, and not against memory. That single discipline is what makes ninety frames of one product survivable.
How do you keep a product looking identical across AI images?
Build a master plate first, generate everything from it, and run a silhouette overlay on every output. Any shift in outline kills the frame and it goes back to the plate.