FILM

Google Research has released FILM: Frame Interpolation for Large Motion, a Tensorflow 2 implementation of a high quality frame interpolation neural network. This approach does not use additional pre-trained networks, such as optical flow or depth, and yet achieves state-of-the-art results. It uses a multi-scale feature extractor that shares convolution weights across the scales for training purposes. FILM can be run with an API on Replicate for research and other applications. This work is jointly authored by Google Research scientists Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru and Brian Curless and is published as a technical report in 2022. If you find this implementation useful, please cite their work accordingly.