Motion Control

Key Insight

An image-to-video model trained on raw clips gives you no say over how much things move — some outputs barely twitch, others thrash. A motion score (also called a motion bucket) fixes this by feeding the model a single number at training time that measures how much motion each training clip actually contains — typically derived from optical-flow magnitude, i.e. how far pixels travel between frames — so that at inference you can dial that number up or down to request subtle or energetic motion. This project adds that input to the Tiny I2V model and checks that the model truly learned the association: low scores should produce gentle, animated-still motion, high scores dramatic movement. It is the simplest control surface for video — one extra knob that separates how much it moves from what is in it.