A quality boundary that moves too slowly becomes perceptible
Aliases: foveal island lag · gaze-contingent update delay · saccadic catch-up
What it is
The full-resolution island in foveated rendering has to follow saccades. Between the eye starting to move and high-quality pixels existing at the new location sit tracking, transport, rendering, and display. If that update latency outlasts the window in which vision resumes encoding, people land on a patch of blur that has not yet caught up, or see the quality boundary drag behind gaze like a curtain. The issue is not how far the periphery was reduced; it is whether the boundary is in place when the next fixation begins.
Why it happens
A typical saccade lasts tens of milliseconds; saccadic suppression takes tens more after landing to lift fully. If the high-quality island is still at the old location after suppression lifts, the new fixation first encodes low-quality pixels, then watches quality “run in.” Albert and colleagues measured this in a headset: around 50–80 ms of delay, observers began to report artifacts reliably, and the threshold fell as the degradation gradient steepened. A steeper peripheral cut is a narrower error budget; the same delay becomes easier to see.
Predictive compensation (moving the island toward the expected landing from saccade direction) can buy time back, but a wrong prediction paints high quality in the wrong place, which is more conspicuous than being late. A permanently larger island also hides delay, at the cost of less saved compute.
Studying it
Make end-to-end delay an independent variable by inserting a known buffer between eye samples and render submit. Have people saccade to a new target and report a blur flash, a sliding quality boundary, or a soft landing point. Also record saccade amplitude: large saccades give the system more warning and make prediction harder. Desktop simulations rarely reproduce a headset’s photon time and warp; re-measure on the target headset. Do not test delay only with smooth pursuit: pursuit is far slower than a saccade and underestimates the problem badly.
Where it stops holding
Apps in which people almost never saccade and only pursue slowly (some viewing content) tolerate more delay. If peripheral quality is reduced only gently, a late boundary is also hard to see—but then little was saved. Dropped samples and blink holes stop or jump the island and look like delay; they must be diagnosed separately from true delay. Some migraine and visual-stress sensitive users cannot continue at delays the lab calls “just noticeable.”
Applying it
- Write eye-to-photon delay into the rendering budget ahead of extra effect quality; saccade-heavy apps get a stricter threshold.
- Hide delay that cannot be squeezed with a moderately larger island or short-horizon prediction, and snap back to the current sample when prediction fails.
- Verify on the target headset with large saccades to high-contrast targets, asking about blur or a sliding boundary at landing, not only mean frame rate.