Keyframes only
job_6d23fbdf928f4ae5The I-frame extract, correctly rebased.
ffmpeg -i https://cdn.rendobar.com/assets/examples/sample.mp4 -vf "select='eq(pict_type\,I)',setpts=N/FRAME_RATE/TB" -t 5 out.mp4video filter
The FFmpeg select filter
Keeps only frames matching an expression, used for thumbnails and keyframes.
3 of 3 variants measured Measured 2026-08-20
The part that catches people
Selecting frames does not renumber them, so the output inherits sparse timestamps and plays at the wrong rate. Pair it with setpts=N/FRAME_RATE/TB or the result is a video that appears to hang.
Measured, not described
Every row below is a real job. The same 1280x720 five-second H.264 source went through each variant, so the numbers are comparable to each other and to every other filter page on this site. Encode time is the FFmpeg step on its own, separated from the download and the upload.
| Variant | Encode | Output | Cost |
|---|---|---|---|
| Keyframes only | 259 ms | 84.2 KB · 1280x720 | $0.0017 |
| Every 30th frame | 292 ms | 130.9 KB · 1280x720 | $0.0018 |
| One frame per second | 316 ms | 234.8 KB · 1280x720 | $0.0018 |
The exact commands
Every 30th frame
job_990386eebe424176Sampling by frame index.
ffmpeg -i https://cdn.rendobar.com/assets/examples/sample.mp4 -vf "select='not(mod(n\,30))',setpts=N/FRAME_RATE/TB" -t 5 out.mp4One frame per second
job_1e65d79263ae4ba4Deliberately missing the rebase, to show the failure.
ffmpeg -i https://cdn.rendobar.com/assets/examples/sample.mp4 -vf "select='not(mod(n\,30))'" -t 5 out.mp4How these numbers were produced
Each variant was submitted to the Rendobar API as an ffmpeg job against the same public source file, and the response was recorded verbatim. Encode time is the API's own per-step timing for the FFmpeg step. Output size is the byte count of the file the job produced. Cost is what the job was actually billed.
Nothing here was estimated, and the whole set is re-runnable, so if a figure looks wrong it can be checked rather than argued about. The 72 variants behind this filter reference cost $0.193082 in total to measure.
Frequently asked questions
What does the FFmpeg select filter do?
Keeps only frames matching an expression, used for thumbnails and keyframes.
What is the most common mistake with select?
Selecting frames does not renumber them, so the output inherits sparse timestamps and plays at the wrong rate. Pair it with setpts=N/FRAME_RATE/TB or the result is a video that appears to hang.
How long does select take to run?
On a 1280x720 five-second H.264 source, the fastest variant on this page (Keyframes only) spent 259 ms inside FFmpeg. That figure is the encode step alone, measured separately from the download and the upload, so it is the filter's own cost rather than the round trip.
Do I need FFmpeg installed to use select?
Not with Rendobar. Every command on this page ran through the Rendobar API against a URL, with no local FFmpeg install and no server. The job id next to each row is the real job that produced the numbers.
Other filters, measured the same way
scale Resizes video to a target width and height, with optional aspect-preserving arithmetic. crop Cuts a rectangle out of each frame, positioned from the top-left corner. fps Resamples video to a constant frame rate by duplicating or dropping frames. transpose Rotates video by 90 degree steps, optionally flipping as it goes. pad Adds bars around the frame to reach a target size without cropping the picture.
Run select without installing FFmpeg
Every command on this page ran through the API against a URL. No binary, no server.