Saccade

A saccade is a rapid, step-like rotation of the eyes that moves the line of sight from one point of fixation to another. Saccades bring the fovea, the small high-acuity region of the retina, onto a new object or region of interest, and people make several of them every second without being aware of choosing their targets. Vision depends on the information taken in during the fixation pauses between saccades; while the eyes are in flight, perception of the fast-moving retinal image is actively suppressed.

Saccades matter to virtual reality and augmented reality in three ways. They set the speed that eye tracking and display pipelines must keep up with for gaze-contingent techniques such as foveated rendering. The brief loss of visual sensitivity around each saccade (saccadic suppression) gives rendering systems a window in which scene changes can go unnoticed, which researchers have used for imperceptible redirected walking. The reduced acuity that follows a saccade has also been used to lower rendering resolution for a short time after each eye movement.

Reviewed 27 September 2026. Checked every figure, date, author list and publication detail against the cited Scholarpedia review, Binda and Morrone review, Wade and Tatler paper, Salvucci and Goldberg paper, Sun et al. 2018 and Kwak et al. 2024 full texts, Crossref and arXiv abstracts for the other papers, and the SIGGRAPH and Road to VR articles. About review dates.

Characteristics

Saccades are stereotyped. In their review for Scholarpedia, John Findlay and Robin Walker write that saccade duration rises non-linearly with amplitude, from about 20 ms for the smallest movements to over 100 ms for the largest, and that peak velocity rises in a similarly regular way, from around 10 degrees per second for small microsaccades to over 300 degrees per second for large saccades. This fixed relationship between amplitude, duration and peak velocity is called the "main sequence", a term borrowed from astrophysics by A. Terry Bahill, Michael Clark and Lawrence Stark in 1975.

Published figures for saccade speed and frequency vary with the task and the range of amplitudes considered. The following values are quoted in sources used in VR and AR research:

Source	Saccades per second	Duration	Peak velocity
Findlay and Walker (2012)	Several	About 20 ms (smallest) to over 100 ms (largest)	Around 10 degrees per second (microsaccades) to over 300 degrees per second (large saccades)
Sun et al. (2018), citing Bahill et al. (1975)	Several	20-200 ms	Up to 900 degrees per second
Binda and Morrone (2018)	Two to three	Not stated	Not stated
Kwak et al. (2024)	Three to five	Typically less than 50 ms	In excess of 200 degrees per second

A saccade is ballistic: its destination is fixed when it starts. According to Findlay and Walker, new visual information can only influence a saccade if it arrives 70 ms or more before the movement begins, although the trajectory shows slight curvature toward the target, which indicates small online corrections. Saccadic accuracy is moderate, at 5 to 10 percent of the saccade amplitude, so a main saccade is often followed after a very short interval by a smaller corrective saccade. Because the motion is stereotyped, its velocity profile and landing position can often be predicted while the eye is still moving.

The delay between the appearance of a new target and the start of the saccade toward it (saccadic latency) varies between people and between trials, but a typical value is around 200 ms. Some latencies are much shorter: so-called express saccades, first reported in monkeys, have latencies of about 80 to 120 ms. Saccades made deliberately away from a target (anti-saccades) take longer, around 250 to 350 ms, compared with a mean of about 150 to 250 ms for saccades made toward a target.

Even during steady fixation the eyes are not still. They drift and tremble and make miniature saccades of less than 0.5 degrees called microsaccades, typically at a rate of around one per second.

Saccadic suppression

The retinal image sweeps across the retina during every saccade, yet this motion goes unnoticed, while comparable wide-field motion produced externally is clearly visible. Paola Binda and Maria Concetta Morrone's 2018 review describes this suppression as selective rather than a general blackout: brief, low spatial frequency stimuli that vary in luminance are very hard to detect when flashed just before a saccade, while high spatial frequency stimuli stay visible and stimuli that vary only in color are not suppressed, or only minimally. The authors conclude that the magnocellular pathway is specifically suppressed, while the parvocellular pathway, which is essential for color discrimination, is left unimpaired.

The time course is short. In the same review, sensitivity to low spatial frequency luminance stimuli begins to fall 25 ms before saccade onset, reaches its minimum at onset, and recovers to normal levels 50 ms afterward. Sun et al., drawing on earlier perception studies, state that suppression occurs before, during and after each saccade and lasts for about 100 ms after the saccade begins. One documented form of the effect is saccadic suppression of image displacement: a scene can be shifted or rotated slightly during a saccade without the observer noticing.

Visual quality does not fully return the instant the eye lands. Researchers at Meta's Reality Labs Research measured post-saccadic visual acuity and found that visual acuity stays reduced for several hundred milliseconds after a saccade, with a rapid increase over the first 100 to 200 ms after the eye lands.

History

The French ophthalmologist Louis-Émile Javal first used the word "saccade" for eye movements, in an 1879 article on reading. Nicholas Wade and Benjamin Tatler have shown that, contrary to a common claim in textbooks, Javal did not measure these movements himself. The measurements came from M. Lamare, working in Javal's laboratory, whose finding that the eye makes several saccades along each line of text (about one for every 15 to 18 letters) Javal reported in a footnote, and from Hering, who described his own studies of discontinuous eye movements during reading in 1879. At the time "saccade" was an ordinary French word meaning "jerk" or "twitch", and its wide adoption in English was due largely to a 1916 article by R. Dodge in Psychological Bulletin.

Detecting saccades in eye tracking data

Eye trackers record a stream of gaze samples, and software has to decide which samples belong to fixations and which to saccades. Dario Salvucci and Joseph Goldberg's 2000 taxonomy of these algorithms describes velocity-threshold identification (I-VT) as the simplest: it computes point-to-point velocities and labels each sample as a saccade point if its velocity is above a single threshold and a fixation point otherwise. It works because eye velocities fall into two distinct groups, low velocities (below about 100 degrees per second) for fixations and high velocities (above about 300 degrees per second) for saccades. The authors note that a two-state hidden Markov model method (I-HMM) gives more robust identification than a fixed threshold.

Headset systems use the same principle in real time. The redirected walking system of Sun et al. estimated gaze angular velocity from the two previous gaze samples at the start of each frame and treated any velocity above 180 degrees per second as a saccade in progress or just finished. The authors note that, because of eye tracker and rendering latency, detection generally lags the actual saccade by tens of milliseconds, and that the lag is tolerable because suppression outlasts it. Bolte and Lappe instead used electrooculography (EOG) and modified two eye movement classification algorithms so that saccades could be detected online, fast enough for scene changes to be time-locked to saccade onset.

Applications in VR and AR

Foveated rendering and latency

Foveated rendering draws a high-quality image only where the user is looking and a cheaper image in the periphery. It generally needs accurate, low-latency eye tracking so that the image stays correct during large, fast eye movements such as saccades. In a 2017 study, Rachel Albert and colleagues at NVIDIA tested three foveation techniques and three foveal radii and found that added eye tracking latency of 80 to 150 ms significantly reduced the acceptable amount of foveation, while shorter added latencies of 20 to 40 ms did not. They concluded that a total system latency of 50 to 70 ms could be tolerated. The authors had framed the study around saccadic omission, which could allow these latency requirements to be relaxed. Sun et al. describe the eye-tracking-to-photon latency of VR systems at the time as around 35 ms.

Another approach is to predict where a saccade will land. In 2017, Elena Arabadzhiyska and colleagues derived a model that predicts the landing position of a saccade while it is under way, so that a gaze-contingent renderer can prepare the image for the new fixation point before fixation is established. The quality mismatch during the saccade itself goes unnoticed because of saccadic suppression. They validated the technique across several combinations of display frame rate and eye tracker sampling rate. A 2022 follow-up by Arabadzhiyska, Cara Tursun, Hans-Peter Seidel and Piotr Didyk examined how saccade orientation in 3D space and smooth pursuit eye movement affect landing prediction in VR and AR applications, and proposed a correction method that adapts existing predictors without extensive new data collection.

Saccade-contingent rendering

In a paper presented at SIGGRAPH 2024, Yuna Kwak and colleagues at Meta's Reality Labs proposed a technique that needs only saccade detection rather than precise gaze position. Their renderer lowers image resolution after a detected saccade and raises it again over several hundred milliseconds as post-saccadic acuity recovers. In a controlled experiment with 30 natural images, observers could not tell the saccade-contingent reduced-resolution rendering from full-resolution rendering. The authors estimate that the method can save more than 50 percent of bandwidth on displays above 60 pixels per degree, and they validated it on a purpose-built eye-tracked headset with 90 pixels per degree and 30 ms eye-to-photon latency.

Redirected walking

Redirected walking lets a user walk through a virtual space larger than the physical room by rotating or shifting the virtual scene slightly as the user moves. Saccadic suppression allows larger shifts to go unnoticed.

Benjamin Bolte and Markus Lappe tested this in a 2015 study published in IEEE Transactions on Visualization and Computer Graphics. They found that participants could not detect translations of about plus or minus 0.5 m along the line of gaze or rotations of about plus or minus 5 degrees in the transverse plane when these were applied during saccades with an amplitude of 15 degrees. They noted that the technique works while the user stands still and suggested it could improve redirected walking.

In 2018, Qi Sun of Stony Brook University and co-authors from Stony Brook, NVIDIA and Adobe Research presented "Towards Virtual Reality Infinite Walking: Dynamic Saccadic Redirection" at SIGGRAPH 2018 in Vancouver. The system ran on an HTC Vive fitted with SMI eye tracking. In a six-person pilot study, no participant detected camera rotation below 12.6 degrees per second (0.14 degrees per frame at 90 frames per second) while their gaze velocity was above 180 degrees per second. The system combined these saccadic rotations with conventional head-rotation gains, used real-time GPU path planning to steer users away from walls, furniture and other people in the room, and added subtle peripheral stimuli to encourage more saccades. The paper's demonstration mapped a 6.4 m by 6.4 m virtual space onto a 3.5 m by 3.5 m physical room, and all users in the study reported that they did not notice the redirection. Anjul Patney and Qi Sun had shown the system earlier at NVIDIA's GTC 2018 conference.

Sun et al. also measured how often saccades would permit redirection in commercial content. In 10-minute gameplay recordings of NVIDIA VR Funhouse and The Brookhaven Experiment, the average share of frames that allowed saccadic redirection was about 11.40 percent and 15.16 percent respectively, enough for angular gains of 1.4 and 1.9 degrees per second.

Eye blinks produce a similar suppression. Eike Langbehn and colleagues showed in 2018 that off-the-shelf eye trackers and head-mounted displays are sufficient to translate a user by about 4 to 9 cm and rotate them by about 2 to 5 degrees during each blink without detection, and estimated that this could improve redirected walking performance by about 50 percent.

Saccade latency and interaction

Saccadic reaction time can itself be affected by what a headset displays. In a 2022 ACM Transactions on Graphics paper, Budmonde Duinkharjav, Qi Sun and colleagues built a probabilistic model that predicts saccadic latency from the statistics of a displayed image and validated it with an eye-tracked VR display. They reported that image changes too small to be seen can still alter how quickly users react, and suggested the model as a tool for gaze-contingent rendering and for fairness in competitive games.

See also

• Eye tracking
• Foveated rendering
• Redirected walking
• Motion-to-photon latency
• Gaze
• Vergence
• Blink