Gyro filtering
Stabilization is only as good as the motion data behind it. A smooth, well-synced gyro track gives a locked-in result; noise, glitches or false rotations in the data turn straight into jitter in the stabilized video. This page covers what makes motion data good and the filters FPVFlow offers to clean it.
What good motion data looks like
Enough samples, recorded cleanly
The camera motion you want to remove is fairly slow compared with what a gyro can measure, so the logging rate does not need to be extreme. What matters more is that each sample is clean.
The main way a sample goes bad is aliasing. A gyro on a quad also picks up motor and prop vibration at hundreds of hertz. If the logger records slower than twice that vibration frequency without filtering it out first, the vibration does not disappear: it folds down into the low frequencies and looks exactly like real camera shake. No filter applied afterwards can tell the two apart. So the filtering has to happen before the data is logged, in the gyro chip or the flight controller. With a camera’s built-in gyro, that part is up to the camera maker.
Gyro and camera moving together
The gyro has to feel the same motion as the camera. With a camera’s built-in gyro this is a given. With Betaflight blackbox data it is not: if the flight controller is soft-mounted and the camera is hard-mounted, or the other way round, the two move differently under vibration and in hard manoeuvres, and the stabilization follows the flight controller instead of the image. Mount them so they move as one body, and keep any play out of the camera mount.
The filters in the Motion data panel
FPVFlow has three general filters in the Motion data panel. Each one is off by default.
Which filters apply depends on the Integration method in the same panel. With None, the stabilization uses the orientation the camera recorded itself (the tooltip reads “Use built-in quaternions instead of IMU data”). With any other method, FPVFlow builds the orientation from the raw gyro and accelerometer samples.
| Filter | Raw gyro and accelerometer | Camera’s built-in orientation |
|---|---|---|
| Low pass filter | Yes | Yes |
| Median filter | Yes | No |
| Glitch filtering | No | Yes |
| Remove DJI Gyrogate false rotations | No | Yes |
All the filters run forward and backward over the clip, so they do not shift the motion in time and do not upset the sync.
Low pass filter
Removes motion faster than the cutoff frequency. The value is in hertz (default 50 Hz when you turn it on), and a lower cutoff means more filtering. It is a second-order Butterworth filter applied in both directions.
Use it when the stabilized video shows fine, fast jitter that the raw footage does not have, which usually means high-frequency noise in the gyro data. Do not go too low: real fast camera motion also gets filtered, and the stabilization stops correcting it.
Median filter
Replaces each sample with the median of its neighbours; the value is the window size in samples (default 5). A median removes isolated spikes, single bad samples far from their neighbours, without blurring the motion around them the way a low pass does. It only works on raw gyro and accelerometer data.
Glitch filtering
Some cameras now and then record a short burst of corrupt orientation data: the rotation rate swings wildly and rings for a moment, unlike real motion, which follows a trend even when it is fast. A low pass would just smear such a burst over its neighbours.
Glitch filtering detects these bursts and replaces them with a smooth path between the last good sample before and the first good sample after. It compares each sample with the local trend and flags the ones that stand out far above the clip’s own normal level, then repeats the search so that smaller glitches hidden next to big ones are also caught.
Strength (0 to 100 %, default 50 %) sets how aggressive the detection is. Higher values catch weaker and longer bursts and search more times, but are more likely to touch real fast motion. Lower values only fix the large, obvious glitches. Regardless of strength, an event that lasts longer than about 0.75 to 2.25 seconds (depending on the strength) is treated as possible real motion and left alone.
Glitch filtering is not in Gyroflow 1.6.3; FPVFlow takes it from newer Gyroflow development code. Gyroflow 1.6.3 and video editor plugins built from it do not apply it.
The low pass filter in Synchronization
The Synchronization panel has its own Low pass filter. It only smooths the data used to find the sync offset, and does not change the motion used for stabilization.
DJI O4 Gyrogate repair
This repair is an FPVFlow feature; it is not in Gyroflow. For the full user guide, see DJI O4 Gyrogate repair.
DJI O4 Air Unit Pro units built since about February 2026, and every O4 Lite, use a gyro chip (TDK ICM-40609-D) that records short false rotations under motor vibration: one axis jumps for a few milliseconds while the other two stay calm. The camera’s own orientation tracking then slowly pulls each false jump back over roughly 12 seconds. The raw video is fine, but the stabilization believes the motion data and “corrects” movements that never happened, so the result jerks and then slowly sways.
None of the general filters can fix this. The false jumps are too large and too fast for a low pass, and the slow pull-back that follows looks like real, gentle motion. FPVFlow therefore has a dedicated repair, Remove DJI Gyrogate false rotations, in the Motion data panel:
- It finds the single-axis bursts in the camera’s orientation track.
- It removes them.
- It adds back the slow pull-back that the camera applied to each removed burst, so the recovery drift disappears too.
- It applies a gentle low pass to the result, stronger in the stretches where a lot was removed.
- Where one axis carries a dense tremor of small single-axis spikes (many per second, below the burst threshold), it low-passes that axis at 8 Hz, fading in and out over about half a second.
The repair runs before glitch filtering and the low pass filter, so they never see the raw bursts. A clip with neither false rotations nor tremor is left completely untouched.
Sensitivity sets how small a jump can be and still count as false:
| Sensitivity | Threshold | Effect |
|---|---|---|
| 0 % | 120°/s | Only the largest, most obvious jolts. |
| 50 % (default) | 60°/s | Tuned against Betaflight blackbox logs. |
| 100 % | 30°/s | Also small jolts, with a higher risk of removing a real knock. |
FPVFlow checks every DJI O4 clip when it opens, reports the likely gyro chip and the number of false rotations, and turns the repair on for affected clips. Save corrected copy writes a copy of the clip with the repaired orientation (<name>_gyrofix.MP4, video and audio unchanged), which other tools, including Gyroflow and the video editor plugins, can then use.
The repair needs the 1 kHz orientation track that the O4 only records with RockSteady and HorizonSteady turned off. Clips recorded with in-camera stabilization on cannot be repaired.