Plain-English explanation
If you have two frames (say the first and the fifth frame of a clip) frame interpolation is what fills in frames two, three, and four so the motion between them looks continuous instead of jumping. It’s a well-established technique from traditional video and animation software, and generative AI models have taken it over for AI-generated video specifically because raw model output is often produced at a lower frame rate than standard video, or with a wider gap between keyframes than looks smooth on its own.
How it works
Older interpolation methods estimate optical flow (the apparent motion of pixels between two frames) and warp pixels along that flow to synthesize the frames in between. Newer AI-based interpolators, often built on the same kind of diffusion or neural-network architecture used elsewhere in video generation, learn to predict plausible in-between frames directly, which tends to hold up better on complex motion, occlusion, or lighting changes that pure optical-flow warping struggles with.
Interpolation can also be used to increase frame rate after generation, turning, for example, a model’s native 8 or 16 frames-per-second output into a smoother 24 or 30fps clip for playback, without re-running the more expensive generation step.
Adult-AI use
Video generation is computationally expensive, so many hosted and local tools generate at a low native frame rate to keep costs and generation time down, then rely on frame interpolation as a cheaper final step to make the result look smooth. For a user, this often shows up as a toggle or a quality setting (“smooth motion,” “interpolate to 24fps”) rather than anything they have to configure by hand.
Common uses
- Smoothing motion in a clip generated at a low native frame rate.
- Increasing a clip’s frame rate for playback without re-running the full generation.
- Connecting two separately generated keyframes with plausible in-between motion.
How it differs from keyframe generation
Frame interpolation assumes the frames it’s connecting are already fixed and correct: its job is purely to fill the gap between them, not to decide what happens. Keyframe generation is the earlier step that decides what those anchor frames look like. A pipeline can use both: generate a handful of keyframes with a video or image model, then interpolate between each pair to produce the full clip.
Limitations
- Struggles with large or complex motion. If the gap between two frames involves a lot of change (object rotation, something entering or leaving frame) interpolation can produce warping, ghosting, or blurred artifacts.
- It can’t fix a bad keyframe. Interpolation only smooths the path between frames; if the anchor frames themselves lack temporal consistency (different face, different lighting), interpolation will blend that inconsistency rather than resolve it.