DAIN

Why we included this project

DAIN is a CVPR 2019 model that creates intermediate frames between two existing ones in a video. It estimates optical flow and uses depth cues to decide which objects should move first, which is what keeps interpolated motion from looking ghosted. Beyond plain frame interpolation, the repo has a slow-motion mode where you set the time step to produce anything from 4x to 100x slowdowns. Pretrained weights ship with it, and there are training scripts if you want to retrain on your own data. The code is research-oriented and dates from the PyTorch 1.0 and CUDA 9 era, so expect to compile custom CUDA extensions and pin an older environment before it runs. That setup effort is the tradeoff for reproducing the published CVPR 2019 results and following the included demos.