FILM
Google's frame interpolator that turns near-duplicate photos into smooth slow-motion and handles large motion better than anything before it.
About
FILM — Frame Interpolation for Large Motion (ECCV 2022, arXiv 2202.04901) — is Google Research's answer to the hardest part of frame interpolation: big displacements between frames, like the ones you get turning two near-duplicate photos into slow motion. The trick is a scale-agnostic motion estimator built on a weight-shared multi-scale feature extractor: large motion at fine scales looks like small motion at coarse scales, so the model gets far more pixels of usable large-motion supervision. A Gram-matrix loss helps it inpaint the wide disocclusions that large motion tears open.
The design is deliberately simple to train — one network, no pretrained optical-flow or depth model, learned from frame triplets alone. Google released the official TensorFlow 2 implementation with pretrained models and a TF Hub module, and it outperformed prior state of the art on the Xiph large-motion benchmark while staying competitive on Vimeo90K, Middlebury, and UCF101 (author-reported). It is one of the most reused interpolation models in open tooling.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (ECCV 2022): SOTA on the Xiph large-motion benchmark; favorable results on Vimeo90K, Middlebury, UCF101
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: Google Research paper (research.google/pubs), ECCV 2022, arXiv 2202.04901; official TF2 implementation and pretrained models via the project site film-net.github.io. Strengths: major-lab backing, genuinely novel large-motion handling, official open release with TF Hub distribution, broad reuse. Docked for: (1) benchmark numbers are author-reported; (2) the googlestaging repo is a static release with minimal ongoing maintenance; (3) TensorFlow-only official code, no PyTorch, no hosted API. License: research code; repo license terms unverified.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ECCV 2022, arXiv 2202.04901 (Google Research: Reda, Kontkanen et al.)
- · Scale-agnostic motion estimator; single network trainable on frame triplets
- · Official TF2 implementation + pretrained models + TF Hub module via film-net.github.io
- · Author-reported SOTA on Xiph large-motion benchmark; static googlestaging release