Video super-resolution aims to create a high-resolution, high-frame-rate video from a low-resolution, low-frame-rate one. Prior methods often lose fine details because they combine frames before aligning them at sub-pixel positions.
This work splats low-resolution features directly onto a high-resolution grid at the correct sub-pixel positions, using learned continuous motion trajectories and visibility weights. This merges temporal interpolation and multi-frame super-resolution into one step.
The approach improves both sharpness and visual quality over current techniques while running up to an order of magnitude faster, because it preserves sub-pixel information throughout the process.