Reproduce “FILM: Frame Interpolation for Large Motion” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, Brian Curless

Google Research developed FILM, a frame interpolation algorithm designed to synthesize high-quality, slow-motion videos from near-duplicate photos characterized by large scene motion. The method achieves state-of-the-art perceptual quality and competitive quantitative benchmark scores on both small and large motion datasets, demonstrating significantly faster inference times.

Key insights

  • Sharing convolution weights across different scales in the feature extractor enables robust generalization from small-motion training data to large-motion scenarios.
  • The application of Gram matrix loss significantly enhances the sharpness and perceptual quality of interpolated frames, especially in challenging disocclusion areas.
  • A streamlined, end-to-end trainable single-network architecture can achieve competitive performance in frame interpolation without requiring auxiliary pre-trained networks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.