Frame Interpolation

FILM: High‑Quality Frame Interpolation from Google Research

A team at Google Research has developed FILM, a single-network frame interpolation model designed to generate smooth motion between two images or video frames. Unlike previous methods, it doesn’t require separate models for optical flow or depth analysis. Instead, it uses a multi-scale convolutional feature extractor with shared weights across scales, delivering top-tier results.

 Technical Highlights

• Unified Architecture: A single neural network handles all frame interpolation tasks, eliminating dependencies on external flow or depth networks.

• Scale-Invariant Feature Extraction: The model processes image triplets at different scales while sharing convolution weights to maintain consistency.

• Minimal Training Data Requirements: Trainable using just frame triplets, reducing data complexity.

From Photos to Video

FILM can take near-identical photos – such as two consecutive frames – and produce slow-motion video that feels naturally fluid, even in complex or fast-moving scenes.

 Accessibility & Integration

• Open-Source Availability: FILM is implemented in TensorFlow 2 and available on GitHub.

• Easy Deployment: It supports Docker and NVIDIA GPU acceleration, making it accessible for local use.

• Online Usage via Replicate: The model runs on the Replicate platform using Python, Node.js, or HTTP APIs. Typical prediction time is under 15 seconds, with costs around $0.0032 per run.

Performance Overview

FILM demonstrates state-of-the-art interpolation accuracy on benchmark datasets like Vimeo‑90K, Middlebury, UCF‑101, and Xiph. It’s capable of generating mid-frame or multiple in-between frames depending on user-specified levels.

Use Cases

• Slow-motion video creation from still images

• Smoothing low frame-rate footage for cinematic quality

• Frame generation for animation or VFX

• Enhancing sports replays and action sequences

• Upgrading video content for modern, high-refresh displays

 Final Thoughts

FILM redefines frame interpolation by combining simplicity and flexibility. It eliminates the need for multiple specialized components, making it easy to use and integrate. Whether on-premise or via cloud APIs, it delivers smooth, high-quality motion suitable for filmmakers, content creators, and research practitioners alike.