Keeping the same person the same person

Decide who they are before you generate anything. Write one page: the face, the hair, the build, every piece of clothing with its colour and whether it is buttoned, and anything they are carrying. Attach the same reference images to every shot and reuse the exact same words. Then check each result against the first shot you were happy with, not against the shot before it. Even with the same prompt and the same references, it drifts.

Usually one of three things. You generated each shot from the one before it, so small errors stacked up. You reworded the prompt somewhere along the way. Or something changed on purpose in one shot, she took the jacket off, and you did not tell the next shot about it, so it guessed.

Inside one short generation, usually yes. Across separate generations and across cuts, no. Faces, clothes, things in hands, sometimes even the whole look of the thing wander often enough that you have to check before you show it to anyone. The benchmarks below measure this and nothing is close to a person yet.

Compare every later appearance with the first clear one, detail by detail. The check here does that across the whole clip in one sitting. It sets a reference for every person, prop and place, checks nineteen named details on each later appearance, from the parting in someone's hair to which hand holds the cup, and shows you the timecode with the two frames side by side.

Each one comes with a note you can use: the shot, the time, what it sees now, what it saw before and where, whether the story explains it, and what to change, naming the frame to match. Copy that into your next generation, or copy the whole lot at once.