google-research/frame-interpolation | Run with an API on Replicate

What

google-research/frame-interpolation is a machine learning model, available through Replicate, that generates intermediate frames between two input images. It is based on FILM, a TensorFlow 2 implementation from Google Research for frame interpolation with large scene motion.

It serves developers, researchers, and media workflow builders who need to create smoother motion transitions or short interpolation videos from two endpoint frames. On Replicate, the core workflow is simple: provide frame1, frame2, and a times_to_interpolate value, then receive either a midpoint frame or an output video with additional interpolated frames. It appears positioned as a research-grade open model that can be used via hosted API or run locally with Cog or Docker.

Features

• Two-frame interpolation input: Accepts a first frame and second frame as required inputs, enabling generation of motion between two still images.
• Adjustable interpolation depth: The times_to_interpolate parameter controls how many times interpolation is applied, which changes output density and practical smoothness.
• Mid-frame or video output behavior: When set to 1, the model returns the sub-frame at t=0.5; when set above 1, it returns an interpolation video with (2^times_to_interpolate + 1) frames at 30 fps.
• API access across common stacks: Replicate provides ready-to-use examples for Node.js, Python, and raw HTTP, which reduces implementation effort for engineering teams.
• Local deployment options: The model can also be run with Cog or Docker, which is useful for teams that want more control over runtime environment and execution location.
• Research-backed model design: The README describes a single-network approach trained from frame triplets and designed for large motion, without requiring separate optical flow or depth networks.

Helpful Tips

• Validate output type early: The page shows both midpoint-frame behavior and video output behavior depending on times_to_interpolate, so downstream systems should be designed to handle the expected file format carefully.
• Benchmark with your own footage: The source describes strong research performance, but production quality will still depend on scene content, motion complexity, and image consistency between the two frames.
• Plan around latency variability: Replicate states that predict time varies significantly based on inputs, so this model may fit better in asynchronous or batch workflows than strict real-time experiences.
• Use local deployment for controlled environments: If data handling, environment control, or repeatable packaging matters, Cog or Docker may be more suitable than only using the hosted API.
• Test interpolation depth conservatively: Higher times_to_interpolate values create more frames, but teams should evaluate whether the extra smoothness justifies longer processing and larger outputs.

OpenClaw Skills

Within an OpenClaw ecosystem, this model could likely support media automation skills that transform sparse visual inputs into smoother motion assets. Likely use cases include an agent that takes two storyboard frames and generates an in-between clip, a content pipeline skill that prepares transition shots for video assembly, or a creative operations workflow that tests multiple interpolation depths and routes outputs for review.

If combined with OpenClaw orchestration, teams in marketing, design, animation, or synthetic media operations could build multi-step workflows around this model, even though the page does not state any native OpenClaw integration. A likely setup would chain frame selection, interpolation, quality checks, asset storage, and metadata tagging into one agentic process, helping professionals turn static visual states into usable motion segments with less manual editing.