Foveated rendering works because your peripheral vision is bad. Outside the fovea, acuity collapses, so rendering the periphery at reduced spatial resolution is nearly free perceptually and very cheap computationally. Headsets with eye tracking have been shipping it for years.
Dropping temporal resolution in the periphery is tempting for the same reason and comes with a problem: judder. Motion at a low frame rate stops looking like motion and starts looking like a sequence of positions.
Masking Judder in Foveated Rendering with Phase-Shifted Noise, posted 7 October 2026 by Ville Cantory, Danhua Zhang, Victoria Interrante and Piotr Didyk, resolves this with a trick that costs almost nothing.
What judder actually is
The paper’s framing is the useful one: judder manifests as aliasing in the spatiotemporal frequency domain.
That reframing is what makes it tractable. A moving edge sampled at too low a temporal rate produces energy at frequencies that were not in the original signal — the same phenomenon as a jagged diagonal line in a low-resolution image, rotated into the time axis. It is not a separate artefact class requiring a separate cure; it is aliasing, and aliasing has known treatments.
The known treatment is temporal filtering — motion blur, essentially. It works, and it carries two costs that matter in a headset: extra computation, which defeats the purpose of foveating in the first place, and a larger loss of detail, because blur removes real signal along with the artefact.
The insight
Prior work established that inexpensive synthesized noise can compensate for the spatial detail and velocity perception lost to foveation. You add procedural noise to the periphery and the visual system reads it as detail it is not actually resolving.
What had not been explored is whether the same noise can restore the motion cues lost to reduced temporal resolution. It can, and the reason is a small piece of mathematics:
Translating a Gabor kernel reduces to a phase shift.
A Gabor kernel is a sinusoid under a Gaussian envelope — the standard primitive for procedural noise with a controlled frequency band, and not coincidentally a decent model of what early visual cortex responds to. Move one spatially and, as far as the sinusoid is concerned, you have changed its phase. So modulating a kernel’s phase creates a perception of smooth motion without having to fully rerender the noise.
That is the whole economy of the method. Motion in the periphery becomes a parameter update on noise you were already generating, not new rendering work.
The part that keeps it from backfiring
Adding moving noise to fix judder could obviously produce a worse artefact than the one you started with — noise that itself aliases, which would be judder with extra steps.
The paper’s actual contribution is the selection rule. Given on-screen angular velocity and the temporal rendering rate, it derives a spatial frequency noise band such that the phase shifts mask judder without introducing motion aliasing of their own.
Both inputs are already available to a renderer. Angular velocity comes from the motion vectors you compute for temporal reprojection; the temporal rate is whatever you chose for that region. So the method is a closed-form band selection over quantities the engine already has — which is why this reads as implementable rather than aspirational.
User experiments show the technique substantially expands the range of angular velocities over which motion at reduced temporal resolution remains smooth, without objectionable degradation of image quality. Eleven pages, thirteen figures.
Why this matters for anyone building immersive work
The frame budget in XR is the binding constraint on everything. Two eyes at 72–120 Hz means every effect, every material, every particle competes for a few milliseconds, and the usual response to running out is to cut content.
Foveation is the main lever that does not cut content — it moves cost to where perception cannot see it. Spatial foveation is mature and widely deployed. Temporal foveation has been held back largely by judder, and it is potentially the larger saving, because halving the peripheral frame rate halves the work for most of the screen area rather than reducing its resolution.
A method that buys back the perceptual cost with a phase parameter on existing noise is the right shape of solution for this constraint: it spends almost nothing and it spends it in the place that was already cheap.
Two things to watch before planning around it. This is a perceptual technique, which means it depends on the periphery staying peripheral — it assumes eye tracking that is accurate and low-latency enough that the foveal region is really where the eye is. And the user study measures smoothness and image quality, not task performance; whether the added noise interferes with peripheral detection of motion you actually need to notice is a different question from whether it looks smooth.