Straight Lines Bend During a Pan. Correct the Clip—or Reshoot the Test?
Straight Lines Bend During a Pan. Correct the Clip—or Reshoot the Test?
A rigid edge leans, or wavers, partway through a pan, and the clip you were going to judge a design on becomes the thing under discussion. The choice in front of you is narrower than it first looks: run a bounded, readout-specific repair against the untreated clip and the question the test was built to answer, or change the capture and admit that the result is a different test. Do both if you can afford it. Keep the source. And treat a smoother result as a smoother result, not as evidence that the camera recorded the geometry it now appears to show.
Identify the distortion without treating every wobble alike
Three different problems present the same way on a monitor: something in the frame looks wrong during motion. They have different causes and different remedies, and mixing them up is how a clip gets processed into uselessness.
The first is whole-frame disturbance. The camera shakes, tilts, or drifts, so the object and the background move together. Within any single frame, the relationship between the panel and the wall behind it is intact; the camera was simply pointing somewhere unexpected. The second is brightness variation: bands or pulses tied to a light source rather than to geometry. That is a separate diagnosis with a separate fix, and it is worth ruling out early because flicker can make an edge look like it is moving when it is only changing value.
The third is the one worth the effort here. The rigid object changes shape inside the frame. Its proportions shift, an edge that should be vertical leans, and so does a straight line in the background that has nothing to do with the object. That is a capture-time problem, not a pointing problem.
The mechanism is bounded and well described in Adobe's public explainer of the rolling shutter effect: many sensors are read line by line, so different regions of a single frame are recorded at different instants. Under fast motion, that difference in instants shows up as skew, which is why a page written to sell software mentions both capture-side and software-side responses. Read it as orientation, not as a fidelity guarantee. A marketing page has an interest in the remedy sounding easy; it does not report a test of whether your particular clip's geometry was recovered.
What that means in practice is a check you can run. Put the camera on a level mount with no roll and no tilt, and pan about the vertical axis only. Under those conditions, every vertical line in the world — the product's edge, a doorframe, a taped stripe on a wall — should appear vertical in every frame. That property survives depth and it survives the pan. If a vertical world edge leans in frame, something is wrong, and if the background leans by the same amount as the object, the cause is shared readout rather than a wobbly product mock-up.
Two shapes of the same error are worth separating. A steady pan produces a uniform lean: every vertical in the frame tips the same way by roughly the same amount and they stay parallel to each other, like a page set in italic. Varying motion during readout — vibration, acceleration, a handheld correction system moving the sensor mid-readout — produces curving edges and the wobble that gets blamed on the operator. The second is much harder to repair with a single setting, and knowing which one you have changes the plan.
A rough proportion helps you decide whether a proposed change is worth making. Take an illustrative case, not a measurement of anything: a 1920×1080 frame with a 60° horizontal field of view, a pan at about 90° per second, and a full-frame readout taking 25 milliseconds. Between the first row and the last, the camera turns 2.25°. Near the center of that frame, one degree is roughly 29 pixels, so the camera's turn moves the whole image about 65 pixels sideways — top row to bottom row. Across 1080 rows that is a lean of about 3.4°. These numbers are arithmetic from assumed inputs, not a sensor characteristic, and the lean is stronger toward the edges of a wide frame than at the center, because the same rotation sweeps more pixels there. What the arithmetic does tell you is the shape of the trade: halving the pan speed halves the lean, and so does halving the readout time. Brand alone tells you neither. The same manufacturer sells sensors with different readout behaviour, and one camera changes character between modes and resolutions. Establish the direction and the rate for your camera and mode, or measure them with the level-mount test.
Most of what follows runs on one concrete plan. Nothing in it has been shot, corrected, or measured; no frame exists and no setting has been changed. It is a comparison being proposed, offered so the checks have something specific to attach to.
The setup: a rigid rectangular mock-up of a proposed product silhouette, matte, with a clean vertical edge, standing on a table. Behind it, a taped vertical line on the wall to one side, positioned so the pan carries the panel across the line partway through the move — the edge occludes the line, then clears it. Camera on sticks with a fluid head, leveled, no tilt, no roll. Pan left to right at a pace that carries the panel through frame in about a second and a half. Same lens and camera position for every version. One locked-off frame from the middle of the arc, shot before anything moves, becomes the reference for the object's true shape.
Preserve what the test was meant to demonstrate
A treatment test exists to answer a question, and the question is rarely "are the verticals straight." It might be whether the product reads as rigid and premium at this pace, whether the silhouette arrives in the order the storyboard needs, or whether a design detail survives the move. Write that down before you touch a slider, because the processing will happily optimize for a different goal.
Three things must exist before any repair happens.
An untouched duplicate of the original clip, copied before anything opens it, with the conditions noted: camera, mode, frame rate, lens, mount, approximate pan speed, whether any stabilization was active. This is the only record of what the camera actually did, and it is the thing you will compare against after the repair has already convinced you it looks better.
A static reference view of the rigid object, shot from the same lens and from a position inside the pan's arc. This is the shape the design is supposed to have. Without it, you have no way to tell whether the object's apparent proportions in the repaired clip match the object or match whatever the correction made convenient.
A frame-matching convention. Know which source frame corresponds to which treated frame. A repair that quietly changes scale or crops the frame makes every later comparison at the same timecode meaningless, and you will not notice unless you established the correspondence first.
The failure mode this prevents is specific. A geometric correction that removes the lean can restore verticals while stretching the object a few percent wider, or pulling the frame in slightly. The clip looks better and the test is now answering a question nobody asked. If the panel ends up 4% wider than the reference, you have not fixed the test. You have moved the error into the one measurement the test was supposed to produce.
The baseline also tells you what "matching" even means here. If the accepted proposal is "this camera, this lens, this pace," then a repaired clip is a proposal that depends on post work the production may not have scheduled. That is a legitimate answer. It just needs to be a stated one.
Try a limited, named repair route
Name the tool, because vague warping is how the escalation starts. After Effects ships a Rolling Shutter Repair effect, and it is the right place to begin: it targets readout specifically, which is the diagnosis you have. Check the labels and defaults against the documentation for the version you actually have installed rather than working from a tutorial written against a different release.
Three controls carry the decision. A rate setting expresses how much readout the effect compensates for, relative to the frame interval. A scan direction setting says which way the rows were read. A method choice selects between a geometric warp and a motion-derived approach such as Pixel Motion. Everything else is refinement.
Do not guess the scan direction. If you correct in the wrong direction you double the lean instead of removing it, which at least makes the mistake obvious: if the verticals get worse, flip the setting. Better, establish it from the camera's documentation, or from careful observation with the level-mount test, and record where the answer came from.
Do not import a percentage from a tutorial as a preset either. The rate that compensates is set by the camera and the recording mode — the sensor's readout time relative to the frame interval — not by the move. How fast the camera was turning during readout determines how much lean there is to remove, not the fraction that removes it. Begin conservatively, increase while watching the background verticals, and stop when they are vertical or when the frame starts to break — whichever comes first. If the background comes right and the object does not, that is information, not a reason to keep pushing. It usually means the motion during readout was not uniform, and no single rate and direction will be correct for every frame.
Choose the method by looking. The geometric warp is the smaller intervention and is usually enough for a steady pan. Move to a motion-derived method only if the warp leaves residuals where the panel's edge crosses the background line. The two do not produce the same result at occlusion boundaries, and the only way to know which suits this clip is to run both over a short, representative range and compare at matched times — not by trusting the method names.
Change one setting at a time, keep the source file closed, and put the result on an adjustment layer or a duplicated composition so a single toggle returns you to the source.
One clarification prevents a common confusion downstream: this repair is not stabilization. It does not lock off the pan, and it should not. Stabilization answers "where was the camera pointing." Readout repair answers "which instant does this row represent." If you stack both, you have changed two things at once and can no longer attribute the result to either. Run them separately first, even if you intend to use them together.
Inspect what the correction displaced
Compare source and repair at matched times, at normal playback speed, across the complete action. A favorable freeze-frame establishes almost nothing about temporal consistency, because the readout error varies from frame to frame and the frame that looks best is often the one where the error was smallest to begin with — or the one where the correction happened to be right by accident.
Walk the same list at several points, including the first and last thirds:
The background verticals. Straight and vertical at each sampled instant, not just in the middle of the move. Lens distortion bends lines near the frame edges, so either apply a lens profile first or restrict the judgment to the central region and accept the edges as a separate problem.
The object's edges. Both vertical, parallel to each other, and free of wobble. If the far edge now leans in counterpoint to the near edge, the correction is overdone.
Proportions. Overlay the static reference silhouette, scaled and positioned to match, and watch whether the object's edges track it through the pan. The object's apparent width will change as the view angle changes; that is perspective, not error. What should not happen is a width that jumps, oscillates, or fails to return when the comparison returns to the same angle.
Occlusion. This is where the plan earns its keep. The interval where the panel's edge crosses the taped line is exactly where a warp that straightens one region tends to smear or tear another. Look at that interval in both directions.
The frame's corners and extremes. Newly bent content, a crop that pushed black borders inward, or stretching at the edges all mean the correction moved pixels that were never misrecorded.
Apparent timing. Does the pan still read at the pace the proposal specified? A correction that alters perceived speed has changed the creative decision while claiming to preserve it.
If fixing one region damages another, reduce the settings or reject the repair. That is a finding, not a failure to try hard enough. And keep the description honest: interpolated pixels are not recovered geometry. A repaired frame shows what you can make the footage look like. It is not a record of what the camera measured, and it should never be filed as one.
Compare a changed capture without pretending it is the same test
If the repair cannot hold across the whole action, change the capture — and label the result correctly. A re-recorded take is not the same test repaired. It is a new test with one condition altered, and the comparison is only meaningful if you say exactly which condition moved.
The lever with the cleanest relationship to the problem is angular speed during readout. A slower pan reduces the lean roughly in proportion, as the arithmetic above suggests, and it costs you whatever the pace was buying: urgency, the sense of a reveal, the rhythm the cut depends on. Say that cost out loud rather than burying it in "we slowed it down slightly."
Other changes are available and each carries its own price. A shorter readout — a different camera, a different mode, or a global-shutter option if one exists in the package — may remove the skew without touching the move, but it changes the hardware the proposal assumed, and that is a production decision, not an editorial one. A different trajectory, such as an arc or a shorter travel, can reduce the time the camera spends turning during readout while keeping the reveal, at the cost of a move nobody storyboarded. Cutting around the fastest part changes the structure of the piece.
What does not fix this is a shutter-speed recipe. Shutter angle governs how much motion blur each row records. It does not change the interval between one row's capture and the next, which is what produces the lean. A very short exposure can actually make the problem more visible, because the blur that was softening the leaning edge is no longer there to hide it. There is no universal setting that settles this for every camera, and the search for one is usually time spent avoiding the comparison.
Where the camera's own guidance covers readout behaviour or recommends modes for motion, use it, and record what it said. Where it does not, the level-mount test from earlier is your measurement: lock the camera, pan at the intended pace, and see how far the verticals move.
One more thing worth stating plainly. If making the test survivable requires changing the move, you have learned something real about the proposal: this camera, at this pace, cannot deliver what the treatment describes. That belongs in the review. It is a result.
File the comparison, not just the winner
The output of this work is not a corrected clip. It is a decision with its evidence attached.
Keep the untreated source, with its conditions recorded and its original file intact. Keep the repaired version with the exact application and version, the source conditions, the settings that were changed, and a note on which frames failed and how. Keep the changed-capture take with an explicit statement of what changed and what it cost creatively. Keep the static reference frame and the result of the vertical-line check.
Give the rejected alternative the same care as the selected one. "The repair held through the first half and bent the background stripe after the panel cleared it, so we kept the source and flagged the pace as unachievable on this camera" is a usable finding. "The repair looked odd" is not.
A decision to leave the clip uncorrected and label the limitation is a legitimate outcome, not an unfinished repair. The repaired clip demonstrates what the footage can be made to look like. The untreated clip and the changed take demonstrate what the camera actually did and what the alternative would cost. A treatment review needs the second pair more than it needs the first.
Frequently asked questions
How do you distinguish rolling shutter from camera shake or flicker?
Camera shake moves object and background together within a frame; flicker changes brightness rather than geometry. Rolling shutter changes shape inside the frame: proportions shift, and vertical edges lean, including background lines unrelated to the object. A level-mount pan about the vertical axis should keep every vertical world line vertical.
What must exist before any repair is attempted?
An untouched duplicate of the original clip with conditions recorded, a static reference view of the rigid object from the same lens, and a frame-matching convention. Without those, a repair that changes scale or crop can make later comparisons meaningless and move error into the measurement the test was meant to produce.
What limited repair route is suggested, and which controls matter?
After Effects' Rolling Shutter Repair targets readout specifically. The decisive controls are rate, scan direction, and method, choosing between geometric warp and a motion-derived approach. Do not guess scan direction or import a tutorial percentage as a preset; start conservatively, watch background verticals, and change one setting at a time.
What should be inspected after correction?
Compare source and repair at matched times at normal playback across the complete action, not from a favorable freeze-frame. Check background verticals, object edges, proportions against the static reference, the occlusion interval where the panel crosses the taped line, frame corners, and apparent timing. Interpolated pixels are not recovered geometry.
What does changing the capture mean, and what does not fix the lean?
A re-recorded take is a new test with one condition changed, so label exactly what moved. A slower pan reduces lean roughly in proportion but costs pace; a shorter readout or global shutter changes the hardware assumption. Shutter speed does not fix the interval between rows' capture, and if the move must change, that is a finding about the proposal.