Open-source eye and colour vision checks that run entirely in your browser.
Colour vision · Visual acuity · Astigmatism · Central field · Perception games
No account. No tracking. Nothing you do here is uploaded.
Try it → · Methodology · Privacy · Contributing
MIT licensed · No dependencies · No build step · Works offline
Important
Fovea is a self-check tool, not a medical diagnosis. It runs on a screen whose brightness and colour profile it cannot measure, in a room it cannot see. It can suggest that something is worth looking at. It cannot confirm or rule out anything, and it does not produce a spectacle prescription.
This is not 100% accurate, and nothing on a web page could be. Published comparisons of screen-based colour plate tests against the printed booklet put sensitivity around 94–96% and specificity around 82–95% — worth doing, nowhere near good enough to decide anything on.
A clear result here is not reassurance — several serious eye conditions cause no symptoms until they are advanced and would not appear in any test on this site. See an eye care professional about any concern with your vision. Full method, limitations and sources: /#/methodology
Most online eye tests are built to capture an email address or sell a product. Almost none of them mention that a browser cannot read your screen's brightness, cannot detect a blue-light filter, and usually has no idea how far away you are sitting — all of which change the answer.
Fovea takes the opposite position. State the conditions, refuse to render a stimulus the screen cannot honestly draw, and say plainly when a result should not be trusted. The honest thing and the differentiating thing turn out to be the same thing.
| Check | What it measures | What it cannot do |
|---|---|---|
| Colour Vision | Red-green and blue-yellow discrimination, plus which type of deficiency the pattern fits | Grade severity precisely, or detect disease that has begun to affect colour |
| Central Field Grid | Distortion, blur or gaps in central vision — where macular problems show first | Detect anything outside the central ~10° |
| Astigmatism Dial | Whether some orientations look sharper than others, and roughly along which axis | Give a cylinder power or a precise axis |
| Visual Acuity | The finest detail each eye resolves, in 6/6 and LogMAR | Measure your actual prescription — only a refraction can |
| Contrast Sensitivity | The faintest contrast you can still detect | Substitute for a chart under controlled luminance |
| Near Vision | Smallest print readable at 40 cm | Replace a reading prescription |
| Red–Green Balance | Whether letters look crisper on red or green | Mean much if you have a red-green deficiency — flagged automatically |
| Hue Arrangement | How cleanly you can order a hue circle | Match a calibrated physical cap test |
| Game | What it shows |
|---|---|
| Find Your Blind Spot | Locates the gap where your optic nerve leaves the retina, and demonstrates your brain filling it in — a line running through the gap stays unbroken |
| Reaction Time | Time from stimulus to response over 7 trials, anticipations discarded |
| Stroop Test | How much a mismatched colour word slows you down. Checks your colour vision result first — this test is meaningless if you cannot separate the inks |
| Peripheral Awareness | How far off-centre you notice something appear, with a central task that verifies you actually held fixation |
Real perceptual measurements, framed as what they are: interesting, shareable, and not a statement about anyone's health.
Colour Vision Simulator — see any image as someone with a deficiency does. Processed in your browser; there is no upload endpoint. 20-20-20 Timer — a break reminder, with an honest note about how well the rule is actually supported.
Every check feeds one printable report: your (optional) name, the date, per-eye findings, and the conditions the test ran under — because a result without its conditions cannot be interpreted. Save as PDF, or export the raw JSON.
The old version of this project shipped ten JPEGs. That was wrong three ways: the answers
were literally the filenames (images/45.webp → "45"), the images were
reverse-image-searchable and memorisable, and clinical plates are not public domain.
Fovea draws every plate on a canvas at run time, with the figure randomised per session.
Twelve plates, all legible. An earlier build shipped 24 across four classes. Two of those classes were removed after rendering them and looking honestly at the result: diagnostic plates carried two overlapping figures and came out as unreadable mush, and hidden plates leaked — the figure stayed partly visible to normal vision, so they measured nothing. A smaller set where every plate is readable beats a longer one where a third are noise.
How the colours are chosen. A dichromat is missing one cone class, which collapses their colour space to two dimensions. The set of colours that collapse to the same point forms a line — a confusion line — and two colours on it are indistinguishable to that person while being obvious to everyone else.
Most implementations derive that line from published copunctal points. Fovea derives it from the null space of the simulation it validates against: the Brettel–Viénot–Mollon transform is a rank-2 linear map in linear RGB, so its null space is exactly the set of directions the deficiency cannot see. Deriving it this way makes the generator and its validator provably consistent. Using published copunctal points instead left a residual separation of up to 0.06 OKLab — meaning "hidden" figures were still faintly visible.
On luminance — a trap worth documenting. Travelling along a confusion line changes brightness as a normal trichromat measures it. It is tempting to conclude the figure could be read from brightness alone, and to mask it with heavy per-dot lightness jitter. That reasoning is wrong and doing it destroys the test: the two pair members simulate to the same colour for the target deficiency, brightness included — measured difference after simulation is 0.0000 for protan. The luminance difference reaches only a normal trichromat, and for them it is part of the signal. An early version scaled jitter to it and quietly erased the figures, worst on protan plates, which have the largest difference. Jitter is now a small fixed mottle for visual character, asserted by test to stay well below the figure/ground gap.
Every plate is validated before display. A plate whose figure is still separable to the deficiency it targets is worse than no plate — it produces a confident wrong answer. The test suite asserts this, plus a minimum legibility floor and a guarantee that the mottle never bridges the figure/ground gap, over 8 randomised sessions.
A run is judged on its worst axis. With plates balanced across protan, deutan and tritan, someone with a strong single-axis deficiency misses only the plates on their own axis — about a third of the set. Their overall ratio then lands mid-band and reads "inconclusive" even though the pattern is unmistakable: every plate on one axis missed, every plate on the others read. Averaging across axes dilutes exactly the signal that matters.
Missing ≥60% of one axis indicates a difference; ≤25% on every axis reads as typical.
There is a deliberate inconclusive band between those. Collapsing it into a binary verdict is where a self-check starts making claims it cannot support.
The demonstration plate is a control, not a question. Everyone reads it regardless of colour vision, so failing it voids the run rather than scoring it.
Verified against simulated observers: a trichromat reads typical, one careless slip still reads typical, and strong protan / deutan / tritan each come out indicated and correctly typed.
A 6/6 optotype at 60 cm is under a millimetre tall, so its stroke lands on well under one physical pixel on a typical laptop. Any site reporting 6/6 under those conditions is measuring its own anti-aliasing.
Fovea computes what your screen and distance can actually render, and asks you to move back rather than reporting a number it cannot draw:
| Display | Distance | Finest honest acuity |
|---|---|---|
| 96 dpi laptop | 60 cm | 6/15 |
| 96 dpi laptop | 2 m | 6/3.8 |
| Retina / phone | 60 cm | 6/7.5 |
| Retina / phone | 1 m | 6/3.8 |
Contrast sensitivity needs dithering to work at all. On a plain 8-bit canvas at mid-grey the smallest contrast step you can draw is about 1.7% — log CS 1.77. Normal peak sensitivity is log CS 2.3–2.7, so every observer with healthy eyes would hit the floor and the test would report "normal" regardless of their vision. It would be measuring the panel. Fovea adds uniform noise of half a code value before rounding, which turns quantisation error into zero-mean noise and lets the spatial average land on the intended contrast. That moves the floor to log CS 2.55 — above normal peak, so the test measures the eye.
The duochrome halves must be luminance-matched. The obvious #FF0000 against #00FF00
is wrong: pure green is 3.4× brighter than pure red, so the observer ends up judging
brightness rather than focus. Fovea uses #009400, whose relative luminance (0.2118) matches
the red's (0.2126) to within 0.0008.
A web page cannot read screen brightness, and cannot detect Night Shift, Night Light, f.lux or an OS colour filter — those are applied after the browser has finished drawing, so reading pixels back reveals nothing. Anything claiming otherwise is measuring its own output. Fovea asks instead, and records your answer alongside the result.
Classic test names are used descriptively where they explain what something is modelled on, never as feature names. Fovea is not affiliated with, endorsed by, or connected to Kanehara Trading, the Isshinkai Foundation, X-Rite, Pantone, Precision Vision, Good-Lite, Richmond Products or Lea-Test. All plates and optotypes here are generated by this software.
Flicker fusion (CFF). A display can only produce a square-wave flicker up to half its refresh rate, so a 60 Hz screen tops out at 30 Hz. Normal CFF is 35–45 Hz. Every observer sits above what the display can present, and the frequencies that are cleanly renderable are only the integer submultiples — on a 144 Hz panel the gaps near 36 Hz are about 7 Hz wide, larger than the entire between-person standard deviation. There is no honest version of this test in a browser.
Stereoacuity. There is no way to send different images to the two eyes on a plain monitor without glasses. Even if there were, one pixel subtends 48–95 arcsec at normal viewing distances — coarser than the 40 arcsec clinical cutoff — so the finest stimulus the screen could present is already worse than "normal". A depth-from-motion demo would work, but it measures a different mechanism entirely and calling it stereopsis would be a lie.
Every figure quoted on the site traces to one of these, and where the research disagreed with itself the more cautious reading was taken. The full annotated list — with a note on why each one mattered — is on the in-app /#/credits page, alongside a visual index of everything that was built.
| Area | Key sources |
|---|---|
| CVD simulation | Brettel, Viénot & Mollon (1997); Viénot et al. (1999); Machado, Oliveira & Fernandes (2009) |
| Colour vision testing | Kanehara plate instructions (38/24/14 editions); screen-vs-booklet validation (2024); weighted-scoring correction (2019) |
| Optotypes & acuity | ISO 8596 (Landolt ring); Bailey & Lovie (1976) logMAR charts; ISO/IEC 7810 ID-1 |
| Contrast & psychophysics | Allard & Faubert (2008) noisy-bit dithering; Campbell & Robson (1968); García-Pérez (1998) staircases |
| Anatomy | Rohrschneider (2004) blind spot position; Ramachandran & Gregory (1991) filling-in |
| Accessibility & safety | WCAG 2.2; MHRA guidance on stand-alone medical device software |
Fovea has nowhere to send your data. There is no server.
This is an architectural fact, not a policy promise:
- No accounts, no cookies, no analytics — not Google Analytics, not a privacy-friendly one, not a self-hosted one.
- No fonts, scripts or images loaded from another company.
- A
Content-Security-Policythat structurally forbids contacting any other origin. Your browser enforces it; we don't have to be trusted. - Results, calibration and preferences live in
localStorageon your device only. - The simulator processes your image in memory and never uploads it.
Verify it yourself: open your browser's network tab and use the whole site. You'll see the initial file loads and nothing else.
The one thing outside our control is the host — GitHub Pages sees what any web host sees when it serves you a file. We never receive those logs. If that matters, clone the repo and open it locally, or install Fovea and use it offline.
An app for people with vision impairments that is itself inaccessible is a self-refuting product. Fovea targets WCAG 2.2 AA for all chrome, navigation, instructions and results.
- Every colour token pairing is contrast-tested in CI, parsed from the stylesheet, so a token cannot be edited without its contrast being checked.
- Status is never carried by hue alone — every badge pairs colour with an icon and a word.
- Full keyboard navigation; a focus ring that survives the mid-grey test field.
- Large-text and high-contrast modes as first-class settings.
- No timed stimuli by default (WCAG 2.2.1), and nothing on the site ever flashes.
One documented exception: the test stimuli themselves deliberately violate contrast minimums, because low contrast is the measurement. That falls under the WCAG "essential" exception, and it is written down rather than left unstated.
No build step, no dependencies, no bundler.
git clone https://github.com/gauravpatil97886/Eye-Health-Color-Blindness-Test.
cd Eye-Health-Color-Blindness-Test.
python3 -m http.server 8080 # or: npx serve, or any static serverThen open http://localhost:8080. file:// will not work — ES modules need an origin.
npm test # colour-science and contrast regression tests (no install needed)The tests need Node 18+ and have no dependencies. package.json exists only so Node treats
the source as ES modules; the site itself is plain files served as-is.
index.html app shell — icon sprite inlined, CSP, no inline script
sw.js offline precache; never touches another origin
assets/
css/
tokens.css design tokens. Type scale steps by 0.1 log units —
the gap between two lines of a LogMAR acuity chart
base.css reset, typography, print stylesheet
components.css buttons, cards, badges, test-surface overrides
app.css screen-level styles
js/
core/ router · store · dom · a11y · prefs · calibration
color/ convert.js (sRGB↔linear↔XYZ↔LMS↔OKLab, WCAG contrast)
cvd.js (Brettel simulation, confusion lines)
plate/ packing.js · glyph.js · generator.js
tests/ registry.js, scoring, and one runner per check
ui/ mosaic.js (the hero), cards.js
views/ one module per screen, lazily imported
test/ contrast + plate-validity regression tests
docs/ design system and research notes
Design principles
- No brand colour on a test surface. A tinted element beside a colour plate shifts its perceived hue through simultaneous contrast. Test surfaces are strictly achromatic and override the theme entirely — enforced by a unit test.
- Nothing animates inside a measurement. A transition between plates would cross-fade two stimuli into a third the observer should never have seen.
- Results carry no good/bad colour axis. A colour vision difference is a difference, not a failure. Red is reserved for system errors and banned from any result.
Corrections are welcome, particularly from anyone with clinical training. The content here is written by a developer working from published sources, not by a clinician, and the About page says so.
Especially valuable:
- Clinical review of the scoring thresholds and result wording
- Translations — Hindi and Marathi first
- Accessibility testing with real assistive technology
- A "kids mode" pass: the plate generator already draws shapes and winding paths instead of numerals, but the surrounding copy is still written for adults
Please don't send screenshots of your results — we don't want them.
Code is MIT. Written content is CC BY-SA 4.0. See LICENSE.md.
Built by Gaurav Patil
Backend engineer · github.com/gauravpatil97886
Fovea began as a college project in 2020 and was rebuilt from scratch in 2026 — new engine, generated plates, twelve checks, and a great deal more honesty about what a browser can and cannot measure.
Full credits and research sources: /#/credits