Make-It-4D: Synthesizing a Consistent Long-Term Dynamic Scene Video from a Single Image

Understand

We study the problem of synthesizing a long-term dynamic video from only a single image.

• This is challenging since it requires consistent visual content movements given large camera motions.
• Existing methods either hallucinate inconsistent perpetual views or struggle with long camera trajectories.
• To address these issues, it is essential to estimate the underlying 4D (including 3D geometry and scene motion) and fill in the occluded regions.