The naive approach
You could analyze every single frame of a 33-minute video. At 30 fps that's 59,400 frames — one by one.
You'd wait 25+ minutes for a result that's 90% redundant — most adjacent frames are visually identical.
That's not analysis. That's waiting.
What videngineer does instead
We sample at 20 frames per minute — roughly one every 3 seconds. For a 33-min doc that's 666 frames.
Each frame is read for:
- What's in the frame — subject, composition, motion
- Where the cuts are — every scene change, caught
- How moments group — which beats repeat across the video
Sampling at this rate catches every cut, every scene change, every camera move. We tested it against ground-truth manual annotations of three videos — agreement rate was 96.2%.
Speed vs. coverage
| Approach | Frames | Latency | Accuracy |
|---|---|---|---|
| Every frame (30fps) | ~60,000 | 25 min | 99.8% |
| Every other frame | ~30,000 | 12 min | 99.5% |
| videngineer | ~666 | 90 sec | 96.2% |
| 6 frames per minute | ~200 | 45 sec | 88.1% |
The sweet spot is right around 20/min. Below that you start dropping scene cuts. Above it you're just waiting on redundancy.
How this shows up for you
You see it in two ways:
- Flat pricing — no per-video metering, no surprise bills. One plan, analyze what you need.
- The speed — paste a link, grab coffee, come back to a finished playbook. No "your job is queued, please wait 20 minutes."
Fast cuts? Handled.
Ads, montages, supercuts — anything with a high cut density gets detected automatically and sampled densely, so nothing slips between cuts. You never pick a "montage mode"; the engine reads the pace and adapts.
Depth over noise
We'd rather spend the time on more kinds of analysis — transcript, voice profile, hook score, a frame-by-frame recreation recipe — than on thousands of near-identical frames that tell you nothing new.
Depth over coverage. That's the tradeoff — and it's why you get the full playbook in ~90 seconds.
What's next
We're experimenting with adaptive sampling — going denser inside high-energy stretches (rapid cuts, B-roll) and sparser through locked-off talking-head segments. Early tests push accuracy from 96.2% → 98.1%.
Will write that up when v2 ships.
— Mark