How does AI find the exercises in a workout recording?
Short answer: The Fit Cut Pro AI Editor splits your recording into five-minute pieces and analyses them in parallel with a multimodal AI model that watches a small, low-frame-rate copy of the footage together with the audio. For each piece it returns a list of intervals: where a set starts and ends, where a rest is, and what the exercise is called. The pieces are stitched back together, overlapping or tiny intervals are cleaned up, and the result lands on a timeline as cues you can rename, move, split or delete before the timers are drawn and the MP4 is exported. In 75 recordings measured between September and October 2026 the median analysis took 95 seconds, and trainers kept 79% of the suggested exercise names unchanged; the whole analysis is saved with the project and can be reset to the original at any time.
How it works
- 1
Probe and prepare
The file is checked for length, resolution, frame rate and audio. Recordings longer than the plan allows (30, 60 or 90 minutes) or shorter than three minutes are turned away before any work starts. A 1080p working copy is made so exports never depend on the size of the original.
- 2
Split into five-minute pieces
Long recordings are cut into five-minute chunks (a short last chunk is merged into the previous one). Each chunk gets a small 10-frames-per-second copy with mono audio: enough to see the movement and hear the trainer.
- 3
Watch and listen
Every chunk is sent to the model at the same time, so a one-hour class does not take much longer than a twenty-minute one. The model looks for the start and end of each exercise, the rests in between and the exercise name. The trainer's own spoken cues (next up, rest, go) are the strongest boundary signal, which is why saying the exercise names out loud helps.
- 4
Merge and clean
The intervals from all chunks are joined on one timeline, clipped to the real length of the video, sorted, and overlaps or sub-second slivers at chunk edges are removed. Each interval is a set, a rest or other footage such as an explanation.
- 5
You correct, then export
The cues appear on the timeline with a preview that matches the export exactly. Rename, drag a boundary, split, mark as no timer or delete; Reset cuts and cues restores the original analysis. On export, timers, exercise names, the next-exercise window and your logo are drawn over the footage in the design you chose.
What we measured (September 1 to October 15, 2026)
Across 75 recordings trainers analysed in the AI Editor, the median recording was 15.5 minutes long and the longest 56 minutes. The median analysis took 95 seconds; nine in ten finished within about 13 minutes, the slow ones being very large 4K files.
The AI found a median of 34 cues per recording, 18 of them exercise sets, with rests and other footage such as explanations making up the rest. Over all 75 recordings that was 1,382 sets, 1,038 rests and 209 other sections.
For 16 recordings we kept the untouched analysis next to what the trainer exported. Trainers kept 79% of the AI exercise and section names exactly as suggested (266 of 335) and changed at least one cue in 9 of the 16 recordings, usually a name or a boundary a second or two off.
Of 98 Editor exports in the same period, 92 finished and 6 failed; the median export took about two and a half minutes on Fit Cut Pro servers.
What the next-exercise window does
In the last ten seconds of a rest, and in the last five seconds of a set that runs straight into another set, a small window shows the next exercise so nobody scrambles. The window shows the demo clip from your library when the cue name matches one of your clips, which is another reason to name clips after the exercise.
What it is good at and what it is not
Continuous class footage with a visible trainer and audible cues: sets and rests are found reliably and most names need no change. Footage with long explanations, several people doing different things, or the trainer out of frame produces cues you will want to check more carefully. The Editor never cuts your footage; it only decides where to draw timers, so a wrong cue is a few seconds of fixing, not a lost recording.
The analysis is free and does not count against your plan; only the export does. Re-running the analysis replaces every cue, including your edits, so fix cues rather than re-analysing unless the first pass was clearly off.
Frequently asked questions
- How accurate is the AI Editor?
- In recordings measured between September 1 and October 15, 2026, trainers kept 79% of the exercise and section names the AI suggested exactly as they were (266 of 335), and changed at least one cue in 9 of 16 recordings before exporting. Boundaries and names are corrected on the timeline in seconds.
- How long does the analysis take?
- The median in 75 recordings was 95 seconds for a median 15.5-minute recording; nine in ten finished within about 13 minutes.
- Does the AI need the exercise names in the audio?
- No, but it helps. It recognises common exercises from the movement; spoken names make rare or custom exercises come out right the first time.
- Is my recording used to train the AI?
- No. The recording is analysed for your project only, and the small analysis copies are deleted when the analysis finishes.
- Why five-minute pieces?
- Short pieces keep each request small and let the whole recording run in parallel; the pieces are merged and cleaned afterwards so cues do not break at the seams.
- Can I see what the AI found before deciding?
- Yes. The cues and preview appear before any export; you spend a workout video only when you export.
- Which plans include it?
- Every plan and the 14-day trial. Recordings up to 30 minutes on the trial and Starter, 60 on Pro, 90 on Ultimate; files up to 25 GB.