A Pattern Crawls in a Test Shot: Change the Pattern, Capture or Review Scale?
A Pattern Crawls in a Test Shot: Change the Pattern, Capture or Review Scale?
The jacket reads as herringbone in your hand and as herringbone with a slow plaid crawling over it on the monitor. The cloth did not change between those two observations. What changed is how often the image asked the cloth a question, and how the answers were added back up.
A crawl like that has two ordinary homes, and they sit far apart. It can be recorded: the sensor sampled a repeat finer than its grid could carry, and what it wrote down is not the pattern. Or it can be manufactured later, when a clean recording is resampled to fit a review window, a proxy, or somebody's laptop. The remedies have almost nothing in common, and one of them costs nothing.
So the first move is not a fix. It is a location. Find the earliest stage at which the pattern becomes unstable, and every decision after that gets cheap.
Separate pattern-related artifacts from other movement
A crawling band is a symptom, not a diagnosis. Four questions will usually sort the aliasing case from everything else that flickers.
Does the disturbance follow the fine repeated detail? Aliasing is locked to the surface. As the cloth turns, the bands turn with it, shifting phase in step with the weave. Sensor noise is locked to the sensor — it stays put in the frame while the subject moves across it. Compression blocking is locked to the codec grid, so it too holds still in the frame while the surface slides underneath. If the artifact moves with the pattern, keep reading. If it moves with the frame, you are looking at something else.
Does it change when the surface gets closer, further away, or turns? Sampling failure is a function of the repeat's projected frequency on the pixel grid, so distance and angle change it immediately and sometimes dramatically. Lighting flicker does not care how close the fabric is.
Does it vanish when the pattern goes soft? Move it slightly out of focus, or step back until the weave is no longer resolved, and aliasing melts away. That behavior is almost its signature.
Does it respond to exposure or ISO? Noise does. This does not.
Everything outside those questions belongs elsewhere. A pulse in overall brightness, a directional shear across the whole frame, a soft haze over the image — different mechanisms, different diagnoses. It is worth saying plainly that a lighting-related flicker is its own problem with its own arithmetic, and no amount of thinking about sampling will resolve it.
The mechanism underneath the aliasing case is worth holding in mind, because it determines what is and is not recoverable. An image is not the scene; it is a set of measurements taken at a grid of positions, each one an average over a small area. Above a certain frequency — the sampling limit, half the grid's density — measurements at that grid cannot distinguish the pattern in front of them from a coarser one. The information is not merely degraded. It is relabelled. Several different source patterns produce identical samples, and once that has happened, the recording no longer contains what you would need to get the original back. Section 8.1, "Sampling Theory," of the fourth edition of Physically Based Rendering, verified on 20 September 2026 is a clean statement of this: samples and display pixels are different things, and a reconstructed image is not the continuously varying source. That is a rendering text, and it supports a mechanism. It does not tell you what is wrong with any particular clip.
One more wrinkle that explains why the worst offenders are prints and weaves rather than smooth gradients. A repeat with hard edges is not one frequency; it is a stack of them. The fundamental and its higher harmonics all arrive at the sensor together. The higher ones are weaker, but they are the ones that fold furthest, and folding far is how a fine weave becomes a coarse, obvious band. A fundamental sitting just past the limit tends to read as a fine shimmer instead — a false texture two or three pixels wide that flips phase whenever the surface moves a fraction of a pixel. That flip is the crawl. Coarse bands and fine shimmer are the same failure wearing different clothes.
Cameras also differ in how much colour detail they sample per position, which is why some moiré arrives in colour — fringing rather than shading. Same cause, different symptom, same stage.
Locate the first stage where the pattern becomes unstable
Before you touch the design or the camera, keep the recording intact. No transcodes, no proxies, no re-exports. The diagnosis depends on measurements you are about to compare, and every unnecessary encode rewrites them.
Establish a known mapping. You need one image pixel to land on one device pixel, and you need to know that you have it. This is harder than it sounds. A viewer's "100%" on a high-density display often means one image pixel per logical point, which can be four physical device pixels. A scaled desktop can resample the whole window behind your back. "Fit" is a resample by definition. So do not trust the label. Verify with a known reference — a single-pixel line or a one-pixel checkerboard placed in the frame — or use an application that reports actual magnification, or export an unresampled crop of known dimensions and present it at a size you can account for. Write down what you established. "The viewer said 100%" is not a record; "one image pixel per device pixel, verified against the chart in frame 1140" is.
Then compare a scaled version of the same passage. Produce it deliberately, at a stated ratio, with a resampler you trust. Compare that against the same passage in the player's own fit-to-window view. If the two behave differently, you have found something. Record the viewer's real ratio — played pixels to image pixels — rather than its zoom setting.
If the disturbance appears only in the reduced version, the processing stage is your suspect. Check the ratio first: integer reductions are usually well behaved; awkward fractions are where cheap scalers show their seams. Check whether the player is quietly working from a lower-resolution proxy, because a 3840-wide source reviewed through a 1920 proxy and then fit to a window is two resamples, not one. Check the scaler itself.
If the disturbance is present in the source at the verified one-to-one mapping, it is recorded, and the rest of this is about what you do next.
Both can be true at once. A recording with marginal detail can alias further when it is reduced. Naming both, and which one you saw where, keeps the fix from drifting into the wrong stage.
Compare pattern redesign with a capture change — on one prop
What follows is a proposed controlled test. No version of it has been recorded, and nothing here is a measured result. It is a protocol designed to produce interpretable comparisons, which is a different and more modest thing.
The prop and the two versions. Take a flat panel 400 mm wide carrying an original repeat you have authored yourself — not a real product's livery. Make two versions of the same design at two pitches, same ink, same substrate, same contrast: Version A at a 1.2 mm repeat, Version B at 0.18 mm.
The geometry. One camera, framed so the panel's full 400 mm width falls across 3840 image pixels. That is 9.6 image pixels per millimetre.
Now the arithmetic that predicts everything. Version A's repeat spans 1.2 × 9.6, about eleven and a half pixels per cycle — comfortable, far above the two-per-cycle floor. Version B's repeat spans 0.18 × 9.6, or just under one and three-quarter pixels per cycle. Below two. It cannot be represented. The seventh harmonic of that 0.18 mm repeat sits near 39 cycles per millimetre, just above four times the sampling rate, and folds to the difference: roughly half a cycle per millimetre, a band about two millimetres wide on the prop, near twenty pixels wide on the recording. Twenty pixels on a frame fitted to a 1920 monitor is ten visible pixels of travelling false pattern. That is the crawl you would expect to see, and it comes from the print, not the monitor.
The move. A slow, repeatable slider or dolly pass, same speed every time. Same lighting, same exposure, same lens, same distance, same codec settings, same review path, same monitor, same player. "Hold everything else steady" is not a formality; it is what makes the comparisons mean anything.
Then three comparisons.
Redesign the pattern. The prediction first, because this comparison is not a clean binary. At this framing Version A's fundamental spans 11.5 pixels per cycle, well clear of the limit, so the coarse travelling band should not be there at all. Version A's own harmonics are a different matter: the sixth and upward cross 4.8 cycles per millimetre and fold, so A can still carry the false texture two or three pixels wide described earlier — the shimmer, weaker as the order rises. That is the wide shot's failure at reduced strength, not a second one.
So swap Version B for Version A and repeat the pass, and judge the result on the coarse band rather than on the absence of any shimmer. If the band has gone, you have located the fault in the recording and confirmed the mechanism, and a residue of fine shimmer in A is what the arithmetic predicts rather than a redesign that failed. The cost is entirely creative and it is easy to underestimate. A 1.2 mm check does not read like a fine twill; it reads heavier, more graphic, less like a garment and more like a graphic. Reducing the contrast of the print flattens the material further. Deliberately breaking the perfect periodicity — making the repeat irregular — kills the coherent beat, but it changes the surface from engineered to handmade, which may be exactly wrong for a package and exactly right for a prop. And if the surface is a real client's own print, none of these routes are yours to take. You cannot redesign a product that already exists.
Change the capture. Keep Version B and tighten the field until only 100 mm of the panel's width falls across the frame's 3840 pixels — 38.4 image pixels per millimetre of panel. Version B's fundamental now spans about seven pixels per cycle, comfortably above the two-per-cycle floor, and the fine shimmer the wide shot produced is gone.
The coarse band is not. Run the fold again at the new density. The seventh harmonic of the 0.18 mm repeat is still at roughly 39 cycles per millimetre, but the sampling rate is now 38.4 cycles per millimetre, so the harmonic sits just past the first multiple of the grid rather than just past the fourth. It folds either way, to about half a cycle per millimetre. That the difference comes out unchanged is a coincidence of these particular numbers rather than a relation carried over: 38.9 − 38.4 and 38.9 − 4 × 9.6 land on the same figure, one harmonic barely above the first multiple of the grid that samples it, the other four multiples above a grid a quarter as dense. Same band, same print, same mechanism. What changed is its pixel scale — about twenty pixels across in the wide shot, about seventy-seven in the tight one, about thirty-eight once that frame is fitted to a 1920 monitor — and a band four times wider is not automatically four times easier to ignore.
If the band itself has to leave the recording, the arithmetic is harsher. The highest harmonic we have been treating as significant — the seventh — has to fall below the new limit, half the sampling rate, which means a density above twice 38.9, or about 78 image pixels per millimetre of panel, and a field about 49 mm wide across 3840 pixels. The cost is that you are no longer showing the panel; you are showing a detail of it. That is a composition decision, and it belongs to whoever owns the shot, not to whoever owns the artifact.
A tempting variant of the capture route deserves a warning. Rotating the panel relative to the sensor does not raise the sample count per cycle at all — the frequency is what it is. What changes is where the beat lands. It can move off a visually dominant axis, land on a diagonal, and become less noticeable, or it can push colour fringing to the fore. Sometimes that is enough. It is not a fix, and calling it one will get you caught in the next setup.
Softening the optical path — a diffusion element, a slight defocus — does genuinely attenuate the scene frequencies above the limit before they are sampled. That is a real anti-aliasing move rather than a cosmetic one. It costs sharpness on everything in the frame, not just the repeat.
Change the review. If the diagnosis pointed at the resample, fix the resample: correct filtering, integer ratios where you can get them, a player you trust, and a re-verification at the scale you will actually deliver at. This route has no creative cost at all, which is exactly why it is worth checking first.
Include a route that fails. After each change, watch the whole move — not a frame. The pass most likely to look successful and be wrong is a detail-suppression pass tuned on a single still: a gentle blur, a denoise, a notch pulled down over the band's frequency. On a still it looks like a cure. Three things go wrong. Scrub the move and the residual instability is still there between frames, because the underlying sampled values have not changed and the filter is simply smoothing over their disagreement. The pass also removes real texture at the same frequencies everywhere in the frame — skin, paper, foliage, the grain of the set — trading a coherent false pattern for a broad loss of true detail. And it leaves you unable to say what the surface actually looks like, which is the one thing the test was supposed to establish.
Retest motion and name what processing cannot establish
Two facts decide this class of problem, and they are easy to reverse.
The first is that recorded aliasing is low-frequency content in the file. The band is a slow, broad pattern — that is the whole reason it is visible. Ordinary downsampling hygiene, which is entirely correct for a review-path fault, preserves low frequencies. It will not remove a recorded band, though it may attenuate it if the band is short relative to the new pixel pitch. So the good habit that fixes the free case does nothing for the expensive one. This is worth saying out loud, because "we'll handle it in the conform" survives mostly on the assumption that the two cases are the same case.
The second is that a filter tuned to the band's frequency is tuned to a frequency that other things in the frame also occupy. Whatever suppression works on the artifact works on them too.
Against that, the case for a post-process is weak but not zero. A restrained pass can reduce a distraction enough that a shot becomes usable, and that is a legitimate production decision. What it cannot do is convert undersampled measurements into an observation of the original surface. Once several different patterns have produced the same samples, no amount of processing can say which one was there.
Watch the whole move after every change, at the scale the piece will actually be seen. A clean still is not evidence of stability. A clean frame at the wrong magnification is not evidence of anything.
And keep the source grounded. The sampling reference supports the mechanism and the reconstruction limit; it does not certify a camera setting, a fabric, an export preset, or the cause of a clip nobody has inspected. The last judgement is a production judgement, made on real footage, in front of the person who owns the look of the shot.
Keep the source and the comparison passage
Preserve the original recording, unaltered, and keep the comparison passage with it — the same seconds of the same move, at the mapping and ratio you established. Keep the failing take alongside the successful one, because it is the only evidence that the route you rejected was actually rejected.
Then write the decision down in three parts, in the order you established them: the first stage at which the pattern became unstable, the route you tested at that stage, and what the route costs. A harder-looking weave, a tighter shot, a softer frame, a rebuilt review path — those are four different prices, and only the production can say which one the treatment can afford.
The useful thing about locating the stage first is that it makes the decision revisitable. If the client later asks for the wide back, or the deliverable moves to a larger screen, or the surface itself changes, you will know which comparison to re-run, and you will still have the passage that shows what it looked like before.
Frequently asked questions
What are the two ordinary places a crawling pattern can originate?
It can be recorded: the sensor sampled a repeat finer than its grid could carry, so the written image is not the pattern. Or it can be manufactured later, when a clean recording is resampled to fit a review window, a proxy, or a laptop. The remedies are different, and one of them costs nothing, so the first move is to locate the earliest stage where the pattern becomes unstable.
Which questions help separate aliasing from other movement artifacts?
Ask whether the disturbance follows the fine repeated detail. Aliasing is locked to the surface and turns with the cloth, while sensor noise stays put in the frame and compression blocking holds to the codec grid. Ask whether it changes with distance or angle; sampling failure does, while lighting flicker does not. Ask whether it vanishes when the pattern goes soft, which aliasing does. Ask whether it responds to exposure or ISO; noise does, aliasing does not.
What does the proposed controlled test compare, and what does it predict?
It compares a pattern redesign with a capture change on one prop. Version A has a 1.2 mm repeat and, at 9.6 image pixels per millimetre, spans about 11.5 pixels per cycle, well above the two-per-cycle floor. Version B has a 0.18 mm repeat and spans just under 1.75 pixels per cycle, below what can be represented. Version B's seventh harmonic, near 39 cycles per millimetre, folds to about half a cycle per millimetre: a band about 2 mm on the prop, roughly 20 pixels in the wide recording, and about 10 visible pixels when fitted to a 1920 monitor.
What happens when you redesign the pattern or tighten the capture?
Swapping Version B for Version A should remove the coarse travelling band, but Version A's own sixth and higher harmonics cross 4.8 cycles per millimetre and fold, so it may still carry a fine shimmer two or three pixels wide. Tightening the field until 100 mm of the panel spans 3840 pixels gives 38.4 pixels per millimetre, removes the fine shimmer, but the coarse band remains; its pixel scale changes from about 20 pixels wide to about 77 in the tight frame, about 38 on a 1920 monitor. Removing the band itself would require a density above about 78 pixels per millimetre and a field about 49 mm wide, at the cost of showing a detail rather than the panel.
What can processing or a review-path fix accomplish, and what can it not?
A review-path fault can be fixed with correct filtering, integer ratios where possible, a trusted player, and re-verification at the delivery scale; that route has no creative cost. Recorded aliasing is low-frequency content in the file, so ordinary downsampling hygiene that helps a review-path fault will not remove a recorded band. A post-process tuned to the band's frequency may reduce a distraction, but it also removes real texture at those frequencies and cannot recover the original surface. Watch the whole move after every change, not a still, and keep the original recording and comparison passage.