October 6, 2026
AI frame interpolation: what it gets right, where it breaks, which tool to use
Interpolation creates the frames missing between two frames of a video, to go from 24 to 30 or 60 fps. Our tests on a fast dance and on a cut between two shots, Topaz against RIFE, the flaws to recognise and the tools, from free to professional.
Frank HoubreFounder of Imaginode
What an interpolation model does
An interpolation model receives two consecutive frames, estimates how each area moves from one to the other, and paints one or more in-between frames; it knows neither the scene nor what really happened between the two moments.
Between two frames of a 24 frames per second video, one twenty-fourth of a second passes. Interpolation means guessing what the scene looked like halfway through that interval, or a third or two thirds of the way, depending on the target frame rate. To do so, the model computes an optical flow: for each small area of the first frame, the direction and distance it moved in the second.
From this it builds a frame where each area has travelled part of the way, then fills the gaps: when an arm passes in front of a wall, part of the wall is revealed, and it has to be painted. That is where models differ. Older algorithms, like ffmpeg's minterpolate filter, work with square blocks and leave torn edges. Neural networks trained for the task, RIFE, FILM, Topaz's own, reason over the whole picture and rebuild revealed areas better.
What the model cannot do is guess an event that appears in neither frame. A blink too fast, a spark, a ball bouncing between two frames: it only sees a before and an after, and draws a straight line between them. For everyday use, taking a shot from 24 to 30 or 60 frames per second, that does not matter. The full conversion method is in convert a 24 fps video to 30 fps.
The hard test: a fast dance
On a dance generated at 24 frames per second, fast arms and motion blur, Topaz and RIFE produced clean in-between frames: no doubled limb, no warped hand, the source's blur hiding the rest.
We expected the hands to break. We took a dance video from our demo catalogue, 720 × 1280, 5 seconds, 24 frames per second, and ran it through two engines on 6 October 2026: Topaz, the one behind the Enhance node, to 30 and 60 frames per second, and RIFE, the most widespread open model, which doubles the frame rate to 48. Then we found, by computation, the two consecutive frames where the picture changes most, and looked at what each engine placed between them.
The result is better than expected. In the crop below, the hand moves from the hip forward; the RIFE frame and the Topaz frame place it halfway, with a consistent blur, no sixth finger, no ghost arm. Part of the reason is the source: a video model generates motion blur on fast gestures, like a real camera, and that blur makes interpolation errors invisible.
One difference showed elsewhere: RIFE, on fal, returns a video without sound, whereas Topaz copies the source audio track. On a silent shot, no matter; on an ad with its music, you have to put the sound back in the edit. Topaz also delivered exactly the requested frame rate, where RIFE can only multiply: 24 becomes 48, 72 or 96, never 25 or 30.
Where it really breaks: the cut between two shots
On a hard cut, interpolation builds a frame that does not exist: Topaz overlaid the two shots in a dissolve, RIFE without scene detection produced a pixel mosaic, and only RIFE with scene detection left the cut intact.
The textbook case is the edit. We joined two shots with a hard cut, two seconds of a village flyover then two seconds of a perfume bottle, and ran the file through the same engines. Between the last frame of the village and the first of the bottle, an interpolation model looks for a movement that does not exist, and still has to produce a frame.
Topaz returned a frame where the bottle shows through the village, like a one-frame dissolve, one sixtieth of a second. RIFE, with no particular option, returned a much worse frame: the bottle cut into squares floating on a background of pixels. With scene change detection switched on, RIFE recognised the cut and left it clean, with no invented frame.
At normal speed, a one-sixtieth of a second ghost frame often goes unnoticed. In a fast edit, with a cut every two seconds, it gives a feeling of dirtiness that is hard to pin down, and it jumps out frame by frame. The rule is simple: interpolate each shot separately, before editing, never the final video.
Other flaws to recognise
Beyond cuts, interpolation gets confused by objects crossing each other, repeating patterns, scrolling text, motion larger than an object's own width per frame, and elements that appear all at once.
Objects crossing, two people passing each other, a hand passing in front of a face: the model must decide which area belongs to which object, and the boundary between them can ripple. Repeating patterns, a grid, stripes, a fence in motion, fool it on correspondence: one bar is mistaken for its neighbour and the pattern seems to slide the wrong way.
Scrolling text and interfaces are another trap: letters warp slightly in the created frames, which shows as soon as you read. Very fast motion, an object moving more than its own width from one frame to the next, has no reliable match: the model may double the object or fade it in. Finally, anything that pops up between frames, lightning, a flash, a spark, is smeared into a dissolve instead of appearing sharply.
In AI videos, many of these cases are rare: models generate little readable scrolling text, and their motion is usually smooth. Pans, drones, faces, rotating products, walking characters all go through very well. They are also the shots where a low frame rate shows most.
The tools, from free to professional
Tools fall into three families: installed software (Topaz Video, DaVinci Resolve, Premiere Pro, Flowframes), online services (Topaz for Web, CapCut, TensorPix, HitPaw) and APIs (Runway, fal); prices checked on 6 October 2026 range from zero to 299 dollars a year.
Topaz Video, formerly Video AI, is the paid benchmark. It offers five interpolation models: Apollo for non-linear motion and slightly blurry sources, Apollo Fast, Chronos for frame rate conversion and slow motion, Chronos Fast for large movements, and Aion for high precision. It costs 299 dollars a year or 59 a month, 699 a year for commercial use above one million in revenue. Its web version, Topaz for Web, targets 30, 60, 90 or 120 frames per second for 19 dollars a month on sale. Adobe signed an agreement in June 2026 to acquire Topaz Labs.
On the free side, Flowframes is a Windows interface for RIFE, DAIN and FLAVR, open source, whose newest builds come out first for supporters on Patreon. RIFE and FILM, the latter published by Google, are research models presented in 2022 and usable from the command line, and ffmpeg offers the minterpolate filter, without a neural network. DaVinci Resolve does optical flow interpolation in its free version; the Speed Warp mode is reserved for the Studio version, at 295 dollars. Premiere Pro, at 22.99 dollars a month, offers Optical Flow in time interpolation. And SVP, at 24.99 dollars for life, interpolates in real time during playback, with no file to export.
Online, CapCut offers optical flow at 30, 50 or 60 frames per second; users have reported its smooth slow motion moving to the Pro plan. TensorPix claims up to 120 frames per second, with a free three-minute trial at 720p, and HitPaw a conversion to 60 or 120, from 24.99 dollars a month. For developers, Runway's API converts to 24, 25, 30, 48, 50, 60 or 120 frames per second, and their 23.98, 29.97 and 59.94 variants, for 0.005 dollar per second. In Imaginode, the Enhance node runs on Topaz: 6 credits per five seconds at 720p up to 30 frames per second, sound kept.
TV motion smoothing
TV motion smoothing is real-time interpolation of 24 frames per second films; it gives films a video look, to the point that Tom Cruise recorded a message in December 2018 asking viewers to switch it off, and a Filmmaker Mode now turns it off by default.
Most recent TVs interpolate everything they show by default, to fill their 60 or 120 hertz panel. On sport, that is progress. On a film, it is what is called the soap opera effect: motion becomes so smooth that the scene looks shot on video, like a TV soap, and loses the slight judder of 24 frames that signals cinema. On 4 December 2018, Tom Cruise and Christopher McQuarrie posted a short video explaining that this setting makes most films look like high-speed video, and how to turn it off.
The UHD Alliance responded with Filmmaker Mode, announced in August 2019 and launched at CES in January 2020, which turns off motion smoothing, noise reduction and sharpening, and respects the original frame rate. On the delivery side, Netflix requires files delivered at their native frame rate, that of the shoot and the edit, and warns that improper frame rate conversions, which leave artefacts, are rejected.
For a creator, the lesson is twofold. Do not interpolate a fiction shot by reflex when it must keep its cinema texture, and choose the frame rate based on where the video will be watched. The choice between 24, 25, 30 and 60 frames per second, platform by platform, is covered in 24, 25, 30 or 60 fps: which frame rate to choose.
The summary
AI interpolation is reliable on continuous motion, pans, drones, characters, even fast ones when the source has motion blur; it gets cuts, text, repeating patterns and instantaneous events wrong.
Interpolate each shot before editing, never a video that is already cut. If you only have the final file, use a tool that detects scene changes, and check every cut frame by frame.
For an exact frame rate, 25, 30, 50 or 60, you need an engine that targets a frame rate, like Topaz; RIFE can only multiply by two, three or four. For a shot with sound, pick an engine that copies it, or plan to put it back.
In Imaginode, the Enhance node changes the frame rate with Topaz at ×1, sound kept, for 6 credits per five seconds at 720p up to 30 frames per second. The ready-made workflow is the "Change a video's frame rate" template.