What the model is computing
Naive frame interpolation averages two neighboring frames. That produces a double exposure: two ghosted copies of anything that moved, which is worse than the judder it was meant to fix.
Real frame interpolation solves for motion first. The model estimates a dense field describing where each pixel in frame A went in frame B, then warps both frames halfway along that field and merges the results, filling any pixels neither frame can explain. This is why interpolation quality tracks motion estimation quality almost exactly: when the motion field is right the in-between frame looks photographed, and when it is wrong you get the melted, rubbery artifacts everyone recognizes.
Three jobs that all use the same tool
- Frame rate conversion. 24fps to 60fps for a smoother screen, or 25 to 24 for a delivery spec. Duration stays fixed.
- Slow motion. Retime a clip you did not shoot at high speed. Duration stretches, frame rate stays. A 4x slowdown of 24fps footage means the model invents three of every four frames.
- Keyframe interpolation. Two fixed images, everything between them synthesized. This is what a first-and-last-frame video model does, and it is video frame interpolation applied over a much larger gap than one frame's worth of motion.
The third case is the hardest by an order of magnitude. Interpolating between adjacent frames means guessing across a 40ms gap. Interpolating between two stills that are seconds apart means inventing a plausible action, which is generation, not reconstruction.
Where interpolation breaks
| Situation | What you see |
|---|---|
| Motion larger than the model can track | Warping, smearing, limbs that stretch |
| Occlusion and disocclusion | Halos and ghosts at object edges |
| Thin repeating structures | Fences and railings that shimmer or bend |
| Text and logos | Wobbling letterforms, the fastest tell |
| Hard cuts inside the clip | A morph across the cut instead of a cut |
That last one is worth checking before every run. If a cut is inside the range you interpolate, the model treats two unrelated images as consecutive and produces a smeared transition. Split the timeline at every cut first and interpolate each shot separately.
Interpolation in an AI video workflow
Generated video makes interpolation more useful than it was with camera footage, for two reasons.
First, cost. Many video models are priced per frame or per second, so generating at a lower frame rate and interpolating afterward is a real saving. A 12fps generation taken to 48fps costs a quarter of what a native 48fps run would, and on smooth camera moves the result is close to indistinguishable. On fast subject motion it is not, so test on your hardest shot rather than your easiest one.
Second, motion cadence. Some models produce clips whose motion feels stepped even at nominal 24fps, because their effective temporal sampling is lower than the container's frame rate. A 2x pass smooths the cadence without touching content.
A practical order of operations: generate, then interpolate, then upscale, then grade. Interpolating after upscaling means the model estimates motion on a frame full of invented detail, and it means paying for a much larger computation for no gain in accuracy.
Reading the result
Judge interpolation by scrubbing, not by watching. Play at quarter speed and watch three regions: the fastest moving object, any edge where one object passes in front of another, and any text. If those three hold, the rest will. If they do not, drop to a lower multiple, split the clip at its cuts, or accept the original cadence, which almost always looks better than a failed interpolation.