Image-to-Video: how to stop i2v from 'releasing' your start frame into a completely different scene
The classic i2v failure: you feed a gorgeous start frame, the first second is perfect, and then the model 'releases' and drifts into a related-but-different scene — the room changes, the character restyles, the color shifts. 2.5 reduced this a lot but it still happens on ambitious clips. Here's how I keep the start frame honoring itself.
LOW motion intent. The more motion you request, the more license the model takes to reinvent. If you prompt 'she turns her head slowly,' it stays home. If you prompt 'dynamic action sequence,' it treats your start frame as a loose suggestion and leaves. Motion magnitude = drift license.
DESCRIBE the start frame in the prompt. Don't just feed the image and prompt the ACTION — also describe the SCENE that's already in the image. Redundant? Yes. But it anchors the model: the image says 'red kitchen' and the prompt ALSO says 'red kitchen,' so it has two votes to stay put instead of one.
Short clips. Drift compounds over time. A 4-second clip drifts a quarter as much as a 16-second one. Generate short, chain if you need length.
The mental model: the start frame is a LEASH, and motion magnitude is how much slack you give. Ask for calm motion and a described scene and the leash stays short.