Jitter: cut, not smoothed over

A short, sharp spike — a marker mis-tracked for a frame or two — is not smoothing work. Post cuts those frames out, peak plus a small pad, and refills the hole with the gapfill model, the way you would replace a spike by hand rather than average it into its neighbours. Jitter that is not a spike is still smoothed.

The switch is Cut and refill jitter on the Run screen, badged · whole run. It is on by default. Its own words:

Impulsive jitter frames are cut out and refilled by the gapfill model before smoothing. Turn off if the result feels too aggressive — the take is then cleaned exactly as before.

Sensitivity

Jitter sensitivity decides how small a jitter has to be before Post leaves it alone. The control's own words:

The acceleration floor under which a jitter is left alone. Changing it re-diagnoses the take — findings from two bands are not comparable.

Band	Floor	What it does
Aggressive	1 g	Flags and repairs the smallest jitter; may touch motion that was fine.
Standard	3 g	Judged in the context of the motion around it. The default, and what a take is diagnosed and cleaned at unless someone chooses otherwise.
Impact	10 g	For scenes with hits, falls and props: only large, isolated spikes are flagged.

Cutting has a higher bar than flagging. A peak can be flagged without being cut: the cut floor is 10 g at Aggressive and Standard, 25 g at Impact. A region card on After says which floor the peak had to clear — Cut floor 10 g at Standard.

Per project and per take

• Per project. Set a default in the project's settings. There is also a No project default row, which is a real choice and not the same as Standard: with no default, a take that says nothing runs at whatever band the workers apply.
• Per take. On Run in Assisted, the control is badged · this take and overrides the project default for that one take.

The one thing it moves

The sensitivity moves the jitter rules and nothing else. Gap fills, swap and merge repair, joins and the fill checks are the same at every band. Within the jitter rules it moves the floor — how hard a frame has to be before it counts — not the shape of the test.

Because findings from two bands are not comparable, choosing a band a take was not diagnosed at re-diagnoses it first. The control says so before you launch:

Re-diagnoses this take — it was diagnosed at Standard, and findings from two sensitivities are not comparable.

You can always see which band produced what you are looking at: Diagnosed at Standard on Before, Cleaned at Standard on After. A take that does not record the band says an unknown sensitivity rather than guessing.

What Post will not do

Three rules that do not move with the sensitivity, or with anything else.

It never creates a spike it did not find

A candidate repair that would put an acceleration jump where your raw capture had none is rejected, whatever else recommends it. The run reports the count of frames it made hot that the capture had not — and that number is meant to be zero.

A repair never exceeds a body's physical acceleration ceiling

Every frame Post writes is checked against an acceleration ceiling for that marker's class — core, limb, extremity, finger — because a torso, an elbow and a fingertip cannot accelerate equally hard. A fill that breaks its ceiling is not shipped: the smoothing steps down through gentler settings until one clears, and if nothing clears, the measured frames are handed back and the hole stays open. A region card reading Nothing filled — the fill moved the marker faster than this marker can move is this rule doing its job.

Smoothing may leave a measured spike that was already over its ceiling over its ceiling. What it may never do is make that frame hotter.

Real fast motion is kept

A strike, a landing, a fall — these are genuine, and the take has them. Before a peak is cut, Post checks whether the marker was actually travelling at that moment. A marker that is moving fast is not jitter, and its frames are left alone. The card states the reading: Flank speed 1.24× its median — not travelling.

One exception, and it is reported rather than hidden: a marker that leaves its trajectory and lands straight back on its own predicted path is cut even when the speed gate declined it. The card says both numbers and the decision, in one sentence: The speed gate declined this peak (1.87× its usual step); it landed 0.4 steps off its own path, so it was cut.

Reading a cut on After

The region card leads with what happened:

• Cut and refilled 6 frames · peak 61 g at frame 1204
• Cut reverted — original frames kept — Post measured the cut and then put the original frames back. Nothing shipped, so there is nothing to revert.

Under it, the evidence: how many measured frames the cut overwrote, how many times the marker's own acceleration level the peak was, the flank-speed reading, the cut floor for the run's band, and whether the cut came from Pass 1 — before the gap fill or Pass 2 — over the filled result.

What to do next

If a take comes back over-cleaned, turn Cut and refill jitter off and run it again — that take is then cleaned exactly as it was before the option existed. If a scene is full of real impacts, move it to Impact instead, and expect the take to be re-diagnosed first. If a specific region is wrong, revert it.

Ask us

Something looks wrong on a take? Write to support@kapnetix.ai and send the take id, or the take's name and the project it sits in.