Four different problems that all look like "it wobbles".
An unstable overlay is usually described one way and caused four different ways: a tremble while the piece is still, a lag while it moves, a jump to the wrong place, or a dropout. They have different fixes, and telling them apart takes about ten seconds once you know what to look for.
Which one are you looking at?
The overlay vibrates slightly even though nothing is moving. Each frame produces a slightly different estimate of where the artwork is, and you are watching that noise directly. Usually means the tracker is working with fewer usable points than it would like — dim light, a sparse design, or the piece small in frame.
The artwork moves and the overlay arrives a moment later, sliding into place. That is lag, not noise, and it is usually the price of stillness — a filter that removes the tremble above necessarily trails during real movement. It also appears when a frame is drawn using a position worked out from an earlier frame.
The overlay is briefly confident and briefly wrong, jumping between positions. That is a matching failure rather than a stability one, and its usual cause is repetition in the artwork: if every point looks like every other point, the wrong match is as plausible as the right one.
Tracking stops and the overlay fades. Glare across the piece, motion blur from a fast pan, or an angle so shallow the artwork is a sliver. The overlay is held briefly and faded rather than cut, because a half-second gap while someone adjusts their grip should not look like a crash.
Why smoothing is not simply turned up.
-
1
Filtering removes tremble
Averaging the position over recent frames cancels the frame-to-frame noise, and a still piece under a still phone looks locked. This is the easy half.
-
2
And the same filter causes lag
A filter cannot tell noise from genuine movement without waiting to see which it was. Turn it up and the overlay visibly trails the artwork whenever anything moves — measured against ground truth, an aggressive filter lags by several times what a light one does.
-
3
So the setting is a choice about your piece
A framed print on a wall that people photograph from a standstill wants more smoothing. A card someone turns over in their hands wants less. There is no value that is correct for both, which is why this is tuned rather than solved.
When no setting is the answer.
Worth saying plainly, because the alternative is an afternoon spent changing numbers that cannot help.
If the piece is small in view, or far away, or the design is sparse, the camera is simply not capturing enough distinguishable detail to fix a position precisely. Every estimate is then uncertain, and no amount of filtering invents the missing information — it only spreads the uncertainty over time.
Move closer, print larger, or add detail to the design. That sounds like a non-answer and it is the real one: the same artwork at twice the size in frame behaves like a different, much better target.
Glare is the other one software cannot argue with. If a laminated card is losing tracking under gallery lighting, the finish is the variable worth changing — see <a href="/posts/printing-for-ar">printing for AR</a>.
Open a template, make it yours.
Interart’s AR templates are finished, working designs — open one in the browser editor, swap in your own artwork, and it’s yours. Each one teaches a technique with a built-in guided tour.
No account needed — sign up later and your work comes with you.
Watch it track something real.
The demo runs the same tracker on a real printed artwork.
Questions, answered.
You are seeing per-frame estimation noise. Each camera frame yields a slightly different answer for where the artwork is, and with few usable points — dim light, a sparse design, the piece small in frame — those answers disagree more. More light and a closer phone both help immediately.
That is smoothing showing its cost. Filtering the position removes tremble while things are still, but a filter cannot distinguish noise from real movement without waiting, so it trails during motion. Less smoothing tracks movement more faithfully and trembles more when still; there is no setting that avoids both.
Almost always repetition in the artwork. Matching decides which point in view is which point on the piece, and in a regular pattern the wrong answer looks as good as the right one. Breaking the repetition, or anchoring on a more irregular part of the design, is the fix.
Not past a point. If the camera has not captured enough distinguishable detail, the position is genuinely uncertain and filtering only redistributes that uncertainty over time. Moving closer or printing larger changes the input, which is the only thing that changes the result.
The overlay is held in its last position for a moment and then faded rather than cut instantly. Brief losses are normal — a hand adjusts, a light catches the surface — and blanking the screen for each one would look far worse than a short hold.