drift

Drift is when AI creates video and the same character or object ends up looking slightly different as scenes change. It's similar to a comic artist who, instead of continually referencing the main character's face drawn in the first panel, relies on memory each time and ends up gradually shifting the position of eyes, nose, and mouth, or the color of clothes, as the panels go on.

This rarely shows up when making a single short image, but the problem stands out when creating longer videos made of multiple connected scenes. If a character's face shape or the furniture arrangement in a room subtly changes from one scene to the next, viewers get a nagging feeling that something is off, even though it's supposed to be the same story.

That's why recent video generation tools try to reduce drift by taking a reference image or video as input and working to keep the people, props, and spatial arrangement in it as consistent as possible.

The article explains that "Alibaba says it reduced this problem by preserving details from reference materials, such as characters, props, and spatial arrangement," introducing how Wan3.0 improved on drift. A common misunderstanding here is that drift is different from making up content that isn't true. Drift doesn't mean the content itself is factually wrong — it refers to a visual consistency problem where the appearance of the same character or object wavers and shifts slightly from scene to scene.

You can attach reference material to a prompt in a video generation tool as shown below and compare whether drift is reduced.

1. Attach a single image that clearly shows the character or object as reference material.
2. Write a prompt that asks for multiple connected scenes based on that image. For example: "Keep the character's face and outfit from the reference image exactly the same, and generate a continuous sequence from the character entering a cafe and sitting down to drinking coffee."
3. Watch the finished video from start to finish and check whether the face shape, clothing color, or background props change from scene to scene.
4. Compare this with a result made using the same prompt but without a reference image — the difference becomes much more noticeable.