Google Research’s Frame Interpolation for Large Motion (FILM) is a high-quality frame interpolation service that’s designed for extensive scene motion. This service has been implemented using TensorFlow 2 and offers a single-network approach that does not rely on additional pre-trained networks, like optical flow or depth, to perform its function. Despite this, it has managed to achieve state-of-the-art results.
A unique characteristic of FILM is its multi-scale feature extractor, which shares the same convolution weights across all scales. Furthermore, the model can be trained from frame triplets alone, making it a standalone solution that doesn’t require any supplementary training or data sources. The FILM project was conceived and executed by Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless at Google Research [1].
The service is available on Replicate, a platform where one can run models with an API, and the model has been run over 122.5K times, suggesting that it is a popular tool among users. Additionally, Replicate provides a comprehensive API reference and examples for users to better understand how to use the service effectively.
FILM’s principal use case is to transform near-duplicate photos into a slow-motion footage. This service interpolates frames, generating new images between two existing images to create an animation. This technology has widespread applications, particularly in video content creation and digital animations.
In the spirit of promoting collaboration and advancement in research, Google Research encourages users who find this implementation useful in their works to cite it appropriately. Details for citation are provided on the Replicate page of the service.
