FILM: Frame Interpolation

FILM is a deep learning-powered frame interpolation model developed by Google Research that transforms static images into fluid slow-motion video by synthesizing intermediate frames. Using a unified single-network approach, it processes two input frames and generates interpolated frames between them, creating smooth motion without relying on external pre-trained networks like optical flow or depth estimation. The model employs a multi-scale feature extractor with shared convolutional weights and is trainable from frame triplets alone. It's particularly effective for large scene motion and can be used through Replicate's API for inference costs around $0.0022 per run. A key limitation is that the prediction time varies significantly based on input characteristics, and the model is archived on GitHub, indicating it may not receive active development.