FILM (Frame Interpolation with Large Motion)

FILM (Frame Interpolation with Large Motion) is a state-of-the-art neural network architecture developed by Google Research, designed to synthesize intermediate frames between two input images. Unlike traditional methods that rely on complex optical flow estimation and separate synthesis stages, FILM utilizes a unified single-network approach based on a multi-scale U-Net architecture. By 2026, it has solidified its position as the industry-standard backbone for open-source video enhancement pipelines, specifically lauded for its ability to handle 'large motion'—scenarios where objects move significantly between frames, which typically causes artifacts in other models. The architecture employs a multi-scale feature extractor that shares weights across different scales, allowing it to capture both fine-grained textures and global movement. This makes it particularly effective for converting low-framerate footage (e.g., 15fps or 24fps) into ultra-smooth 60fps or high-speed slow-motion (x8, x16) without the 'soap opera effect' artifacts. It is widely integrated into VFX workflows, medical imaging transitions, and creative AI tools where temporal consistency is paramount.