Recently went through some early image-to-video approaches, and the more I looked the more the two lines of thinking stood out in contrast. One splits it into two stages: first infer pixel motion trajectories from the static image, then use motion-enhanced temporal attention to synthesize subsequent frames along that trajectory — basically plan first, then paint.
The other goes the spatiotemporal diffusion route, stuffing temporal-dimension downsampling and upsampling directly into the Unet, denoising the whole video as one spatiotemporal volume together — Lumiere is the representative of this camp.
The former is controllable but depends on how accurate the trajectory estimation is, the latter is more holistic but eats up compute. Basically all current models show traces of these two ideas.