IMPROVED VIDEO DIFFUSION WITH NOISE SCHEDULING

Summary from the TLDR AI newsletter

Generating a single image from a single prompt is generally fine, but fails when generating video because of the temporal changes between frames. However, consistency when changing text is quite challenging. This work tackles both of these problems and enables generations up to 512 frames long using diffusion.