Generating videos that follow a user's motion instructions, like a camera pan, is challenging. Current methods alter the random noise that starts the generation process to reflect the motion, but this distorts the noise's special statistical structure, degrading quality and requiring slow corrections.
We introduce Householder Motion Embedding (HME), which applies a series of mathematical rotations (orthogonal transformations) to the noise at each pixel. These rotations encode the motion without harming the noise's natural Gaussian properties, removing the need for corrections.
In benchmarks, HME matches or surpasses existing motion-control methods in accuracy and quality while achieving a 2.7 times speedup in noise processing. It is a plug-and-play upgrade for any video .