Pelli-Robson vs FACT vs qCSF: a tour of the major contrast sensitivity tests

Three families of contrast sensitivity test dominate clinical and research use today: a single-frequency letter chart (Pelli-Robson), a fixed-step grating chart (FACT, with the closely related CSV-1000 and Hamilton-Veale), and adaptive psychophysical procedures (staircases and the Bayesian qCSF). Each was designed for a different purpose, and the numerical outputs are not interchangeable. A "log CS of 1.85" from a Pelli-Robson chart is not the same quantity as a "log CS of 1.85" from a FACT chart row, and neither is directly comparable to an AULCSF returned by qCSF.

This post walks through what each test is, what it actually measures, and where its known limits are. It is written for clinicians evaluating tools, vision researchers picking a method, and methodologically curious readers who want to know which number on a results page is doing what work. For the conceptual background on what contrast sensitivity is, the primer is the better starting point; this post assumes that vocabulary.

Why so many tests exist

Different tests sample the contrast sensitivity function in different ways, with different precision, for different reasons. A bedside screening tool needs to be fast, language-independent, and produce a number with established norms. A research instrument needs to recover the full curve with high precision. A multi-site clinical trial needs reproducibility across sites with chart-based simplicity. A self-tracking app on a consumer device needs to balance precision against time and against the variance that calibration introduces. No single instrument is optimal for all four jobs, and the field has accumulated a small family of tests that each prioritize a different point on those tradeoffs.

The grandparent: Pelli-Robson (1988)

The Pelli-Robson chart was introduced by Denis Pelli, John Robson, and Arnold Wilkins in 1988.¹ It is a single printed chart of Sloan letters, all the same physical size, arranged in triplets — sixteen letters per row, eight rows. At the standard 1 m viewing distance the letters sample contrast sensitivity at roughly 1 cycle per degree of visual angle, near the low-to-mid range of the CSF. Reading down the chart, contrast decreases in steps of 0.15 log units (about a factor of 1.4 in linear contrast) per triplet. The observer reads aloud until they fail two consecutive triplets; threshold is taken at the last triplet where at least two of three letters were named correctly. Letter-by-letter scoring (0.05 log per letter) gives finer resolution than triplet-by-triplet scoring.

The output is a single number: log contrast sensitivity at one spatial frequency. Healthy young adult monocular values are around 1.84 log units at age 20-39, declining to about 1.68 by age 60 and above; binocular sensitivity is roughly 0.15 log units higher.² Test-retest repeatability on the chart itself is approximately ±0.15 log units (about one chart step), with the smallest clinically meaningful change conventionally taken as ±0.30 log units.³ Those repeatability figures come from Elliott, Sanderson and Conkey's 1990 reliability study — not from the original 1988 design paper, a distinction worth keeping straight since the two are often conflated.

Strengths: simplicity, speed (under two minutes), and three decades of accumulated normative data. The limits follow from the design. It samples one spatial frequency; a patient with a mid-band CSF deficit at 4-6 cpd and intact low-frequency sensitivity may score normally. Chart contrast depends on print quality and ambient illumination; a faded chart or a darker room shifts the effective contrast. And the letter task confounds visual sensitivity with letter-recognition variance, which matters for non-native readers and dyslexic observers.

Note: Pelli-Robson values are reported in log contrast sensitivity computed against Weber contrast (target vs. background). Grating-based tests below use Michelson contrast. The two scales are conceptually related but not numerically interchangeable for the same observer.

It remains the most-cited single-frequency contrast sensitivity instrument in clinical work — particularly for longitudinal tracking of cataract, glaucoma, optic neuritis, and amblyopia where a single fast number on a standardized chart is exactly the right tool.

The chart-based grating CSF: FACT, CSV-1000, Hamilton-Veale

The Functional Acuity Contrast Test (FACT, originally Vistech, attributed to Arthur Ginsburg) replaces letters with sine-wave grating patches at five spatial frequencies — 1.5, 3, 6, 12, and 18 cpd — and nine fixed contrast steps per frequency. Each patch is tilted left, right, or vertically; the observer names the tilt of each patch in a row and the row threshold is the last correctly identified patch. The chart returns five log CS values, one per frequency, which together form a five-point CSF.

The closely related CSV-1000 (VectorVision) presents gratings at 3, 6, 12, and 18 cpd in a self-illuminated wall chart maintained at 85 cd/m², using a 2-alternative forced-choice (top vs. bottom) response that reduces the guess rate compared to FACT's 3-alternative tilt task. The CSV-1000 is among the more common contrast-sensitivity outcome measures in clinical and device trials, valued for its standardised self-illuminated luminance. The Hamilton-Veale chart, used widely in optometry, is a printed sine-wave grating chart with a similar discrete-step structure.

Strengths: a five-point CSF in three to five minutes, on hardware that fits on a wall and doesn't require a computer. For practitioners who want more than Pelli-Robson's single number but don't have access to a research-grade system, FACT and CSV-1000 are reasonable workhorses.

The weaknesses follow from the fixed-step design.⁴

1. Coarse contrast quantization. Nine steps at roughly 0.15 log-unit spacing is about a 30% contrast change per step. The smallest detectable change at a given frequency is one full step, at the edge of what's clinically meaningful on more precise instruments.⁵
2. Ceiling effects near the peak of the curve. Because the top of the FACT contrast scale sits below the sensitivity many people with good vision actually have, a large share of observers score the maximum possible value at low-to-mid frequencies — so the chart cannot separate "good" from "great" there, or reliably track improvement in already-normal eyes. In one refractive-surgery cohort, half of post-LASIK eyes hit the maximum FACT score at 1.5 cpd; the authors concluded the chart's compressed range produces a "truncated scale" that fails to discriminate among people with good contrast sensitivity.⁵
3. Single trial per cell. One presentation per (frequency, contrast) combination yields higher test-retest variance than adaptive procedures averaging over many trials near threshold.
4. No adaptive resampling. Trials are wasted on contrasts the observer trivially sees or clearly cannot; the procedure does not concentrate sampling near threshold where information density is highest.
5. Calibration dependence. Printed and self-illuminated charts both depend on maintained illumination level and on the chart not having faded.

None of this makes FACT useless. For a quick coarse map of the CSF in a clinic, it is fine. It is not a precision instrument and was not designed to be one.

Modern psychophysics: the 2-down-1-up staircase

The procedures above are non-adaptive — stimulus levels are fixed in advance. Adaptive psychophysical methods change that: the next contrast presented depends on the observer's response history, and trials concentrate near threshold.

The simplest adaptive method that converges to a defined point on the psychometric function is the transformed up-down staircase described by Harry Levitt (1971).⁶ The rule is procedural: after two consecutive correct responses, decrease contrast (make it harder); after one error, increase contrast. This "2-down-1-up" rule converges to the contrast at which the observer is correct on √0.5 ≈ 70.7% of trials — a point near the steep middle of the psychometric function where threshold estimates are statistically most precise. Other variants converge elsewhere (3-down-1-up to 79.4%, 4-down-1-up to 84.1%, 1-down-1-up to 50%); 2-down-1-up is the workhorse for two-alternative forced-choice contrast tasks because the 70.7% target sits comfortably above the 2AFC chance rate of 50%.

The procedure usually starts with a large contrast step that halves at each reversal, and the threshold estimate is the mean of the last six to eight reversals. Forty to sixty trials per spatial frequency is typical for a 0.1-log-unit threshold estimate, with each frequency run independently.

Strengths: simple to implement, reproducible, statistically well-characterized, and decent precision in a manageable trial count. Weaknesses: each frequency is estimated independently, so the procedure does not exploit the underlying smoothness of the CSF curve; total trials add up linearly with the number of frequencies. For a five-frequency CSF this is roughly 200-300 trials, which is more than qCSF needs.

This is the method our app uses. It is a deliberate engineering choice, not a claim of methodological superiority over qCSF — discussed below.

The Bayesian frontier: qCSF

The quick contrast sensitivity function (qCSF) procedure, introduced by Lesmes, Lu, Baek, and Albright in 2010, replaces independent per-frequency thresholds with Bayesian estimation of the whole CSF curve.⁷ It generalises to the full curve the single-threshold Bayesian-adaptive approach pioneered by Watson and Pelli's QUEST,⁸ and is the current state of the art for measuring the full CSF in a short session.

The core insight is parametric. The CSF across spatial frequencies is well-described by a truncated log-parabola with four parameters: peak sensitivity, peak spatial frequency, bandwidth (full width at half-maximum in octaves), and a low-frequency truncation level. Knowing these four numbers reproduces the curve. The qCSF procedure maintains a posterior probability distribution over that four-parameter space and, before each trial, picks the (contrast, spatial-frequency) combination that maximizes expected information gain about the parameters. Each trial therefore informs the whole curve rather than a single point — a trial at 3 cpd refines the estimated sensitivity at 12 cpd as well.

Reported performance is striking. A usable area-under-log-CSF (AULCSF) summary metric is achievable in about 25 trials (~2 minutes), and the original method paper reported a test-retest correlation of about 0.96.⁷ Stimuli are typically band-pass filtered letters, gratings, or Gabor patches; the response is an N-alternative forced choice on orientation, letter identity, or location.

Strengths: precision per unit time is much higher than fixed-step or per-frequency staircase methods. The output is the whole CSF curve plus interpretable summary metrics (AULCSF, peak gain, peak frequency, cutoff) useful for tracking change over time.

Weaknesses worth naming:

1. Model dependence. qCSF fits the four-parameter truncated log-parabola to the data. Observers whose true CSF deviates from this shape — a notch, a non-smooth dip — will have their data smoothed into the parametric form. Other parametric CSF models exist in the literature; the truncated log-parabola is the Lesmes-Lu-Baek-Albright choice and is what the published qCSF procedure assumes.
2. Implementation complexity. A correct qCSF requires maintaining a four-dimensional posterior, computing expected information gain per candidate trial, and rendering carefully calibrated stimuli at adaptively chosen contrasts and frequencies. Bugs are easy and not obvious in the output.
3. UI surface. A multi-parameter Bayesian procedure produces a richer output but is harder to explain to a non-research user.

For research, multi-site trials, and serious clinical use, qCSF is the right answer when correctly implemented and validated.

Comparison

A practical note for clinicians comparing numbers across instruments: the units overlap but the quantities do not. Pelli-Robson log CS is sensitivity at one frequency on a letter task with Weber contrast; a FACT row threshold is sensitivity at one frequency on a tilt task with Michelson contrast; a qCSF AULCSF is an integral of the fitted log-CSF over a defined band. The right comparison is within-instrument over time on the same observer.

A row of Gabor patches at increasing contrast

For readers less familiar with what a grating contrast sensitivity stimulus actually looks like, the FACT row is a sequence of Gabor-like patches at the same spatial frequency with contrast stepping down from left to right. A simplified version:

high contrast

A FACT row asks the observer to name the tilt of each patch until the patches disappear into the gray; the last correctly identified patch defines the threshold for that frequency. A staircase or qCSF, by contrast, adaptively picks which contrast to show next based on whether the observer got the previous trial right, and never presents obviously-easy or obviously-impossible patches except briefly during convergence.

What we built, and why

VCS-Test runs a 2-down-1-up adaptive staircase on Gabor-patch stimuli at multiple spatial frequencies. The output is a per-frequency log contrast sensitivity curve plotted against age-stratified normative ranges from published cohorts.²

The choice of 2-down-1-up over qCSF for v0 was deliberate, on three considerations. First, the math of a transformed up-down staircase is small, well-characterized,⁶ and verifiable end-to-end; explaining it to a clinician or a curious patient takes a paragraph, not a model derivation. Second, a four-to-five-frequency staircase fits inside a five-to-seven-minute consumer session — long enough for precision, short enough that fatigue and attention drift do not dominate the variance budget. Third, the precision per session is appropriate for a screening signal on consumer hardware where display calibration introduces its own variance floor; spending extra precision on the procedure while accepting calibration noise upstream is wasted effort.

The honest framing: per-frequency precision is meaningfully better than FACT's nine fixed contrast steps; total-trial efficiency is meaningfully worse than qCSF. A future v2 could swap in a qCSF mode for users who want a full curve in fewer trials, particularly once we have published our own test-retest reliability data. Whichever procedure runs, the calibration upstream — credit-card screen sizing, blind-spot viewing-distance estimation, perceptual gamma probe, mid-gray surround — matters more for accuracy than the choice between procedures.

For implementation notes, see the methodology page. For the conceptual primer, what contrast sensitivity actually measures. For a practical walkthrough, how to take a contrast sensitivity test online.

Note: A contrast sensitivity test, however well-implemented, is a screening signal of overall visual function. It is not a diagnostic test for any specific condition, and any single result should be read alongside an in-person eye examination rather than in place of one.

Take the test

Take the test now — our free, calibrated 2-down-1-up adaptive staircase, in your browser, about three minutes for the quick mode. It plots a per-frequency log contrast sensitivity curve against age-stratified norms, and the result stays on your device by default. As with any instrument above, the most informative comparison is your own trend over time on the same setup.