How Accurate Are the Windows Color Filters?

Q: How accurate are the Windows 11 color filters?

Nobody outside Microsoft can say. Microsoft documents that the color filters exist but does not publish the transform they apply, and the filters run after the desktop image is composed — below the level any screenshot, screen recording or public API can observe. Three independent capture paths were tested and all three returned the unfiltered image. This does not mean the filters are ineffective. It means their accuracy cannot be checked from outside.

Q: Why don't Windows color filters show up in screenshots?

Because the filter is applied after the desktop is composed, at the display output stage, rather than to the image Windows hands to capture tools. A screenshot is taken before the filter is applied, so it shows the uncorrected colors. The same is true of screen sharing and screen recording — Teams, Zoom, OBS and remote desktop all transmit the unfiltered image. Night Light behaves the same way.

Q: Can I find out what color transform Windows applies?

Not through any public software interface located so far. Microsoft's own MagGetFullscreenColorEffect API — the documented way to read a full-screen color effect — returns an exact identity matrix while a color filter is active. GDI screen capture and DXGI Desktop Duplication both return byte-identical frames with the filter on and off. Recovering the transform would require measuring light leaving the display with a camera or colorimeter.

Q: Should I use the Windows color filters or a third-party tool?

The Windows filters are free, built in, apply to everything on screen and cost nothing to try — start there and judge by whether they help you. The practical reason to add a separate tool is that a published transform can be inspected, adjusted in strength, and will appear in your screenshots and screen shares, which the built-in filters will not.

Nobody outside Microsoft can say. Windows 11's color filters for protanopia, deuteranopia and tritanopia are applied after the desktop image is composed — below the level any screenshot, screen recording or public API can observe. We tested three independent capture paths and all three returned the unfiltered image, byte for byte. Microsoft doesn't publish the transform, so its accuracy can't be checked from outside. That is not a claim the filters are bad. It's a claim that nobody can tell.

These filters are plausibly the most widely deployed color-vision assistive transform that exists — they ship on every Windows installation, switched on from Settings → Accessibility → Color filters. We went looking for their coefficients in order to compare them fairly against a published daltonization method. We couldn't find them, and that turned out to be the interesting part.

What we tested, and what came back

Three independent software paths, on Windows 11 Pro build 26200:

Results of three capture paths with a Windows color filter active
Path	What uses it	Result
GDI / BitBlt	Ordinary screenshots, Print Screen, Snipping Tool	Blind. 0.0000 of pixels changed
DXGI Desktop Duplication	OBS, Teams and Zoom screen-share, remote desktop	Blind. 0.0000 of pixels changed
`MagGetFullscreenColorEffect`	Microsoft's own documented API for reading a full-screen color effect	Returns identity while a filter is active

Not "small differences." The captured frames were byte-identical while the screen was visibly changed to the person sitting in front of it. The obvious way to get a result like this by accident is to run an experiment that never changed anything, so the filter's state was read out of the registry at the moment of every capture and stamped into the file; the comparison tool refuses to draw any conclusion from two captures that share a state. An earlier run of ours was invalid for exactly that reason.

The part that affects you today

If you use Windows color filters, your screenshots do not show what you see. Neither does your screen share. When you send a colleague a screenshot to ask "can you tell these two lines apart?", they are looking at the uncorrected image — and so is the support agent, the bug tracker, and the person on the other end of the Teams call.

This catches people out constantly, and it isn't a bug in your setup. Night Light behaves the same way for the same reason. If you need to show someone what your screen actually looks like, photograph the monitor with a phone — that is currently the only way.

Why it works this way

The evidence points to the filter being applied in the display output pipeline — the desktop compositor's output stage or the graphics card's output color conversion — rather than to the composed desktop image that capture tools read. That's a sensible place to put it: it's fast, it costs nothing per application, and it genuinely does cover everything on the display, including games and video that other approaches miss. The side effects follow from the same design decision that makes it work well.

What this doesn't mean

Being careful here matters, because the honest finding is narrower than the interesting one:

It doesn't mean the filters are ineffective. They may be excellent. We measured that they are unmeasurable by these means, which is a different statement.
It doesn't mean no interface exists. Three public paths returned nothing. A private or undocumented one may well exist.
It doesn't generalize past what we ran. One machine, one GPU vendor, one Windows build. A different setup could behave differently, and the test is cheap to repeat — the scripts are public.
The screenshot behavior was already known informally. Accessibility guides have noted it for years. What's new here is the measurement, the extent, and the API result.

The one result we haven't found reported anywhere else is the third row of that table: Microsoft's own full-screen color-effect API returns an exact identity matrix while a color filter is running. That's the result that closes off the obvious workaround, and it's why no accessibility checker, testing harness or third-party benchmark can audit these filters.

What we'd rather have

Publishing the three matrices would cost Microsoft very little and would let anyone verify, improve on, or teach the transform. Until then the only remaining route to measuring them is a camera or a colorimeter pointed at the display, which is where this work goes next.

Everything above is reproducible. The method, the controls, the raw captures and the scripts are public and MIT licensed, including results that don't flatter our own correction — several of our candidate matrices were rejected by our own test.

If you're choosing what to use

Start with the built-in filters. They're free, they're already installed, they cover the whole display, and whether they help you is something you can judge in thirty seconds without anyone's benchmark. Reach for something else when you want a transform you can read, a strength control, or corrected colors that survive a screenshot.

Set up color correction — browser, whole screen, or your own site →

Frequently asked questions

How accurate are the Windows 11 color filters?

Nobody outside Microsoft can say. The transform isn't published, and three independent capture paths — including Microsoft's own color-effect API — can't observe it. That's a statement about verifiability, not about quality.

Why don't Windows color filters show up in screenshots?

The filter is applied after the desktop is composed, at the display output stage, so a screenshot is taken before it happens. Screen sharing and recording are affected the same way, as is Night Light.

Can I find out what color transform Windows applies?

Not through any public software interface we've located. MagGetFullscreenColorEffect returns exact identity while a filter is active. Recovering it would need a camera or colorimeter measuring the display itself.

Should I use the Windows color filters or a third-party tool?

Try the built-in ones first — free, installed, whole-display. Add a separate tool when you want an inspectable transform, a strength control, or correction that survives a screenshot.

Measured 4 August 2026 on Windows 11 Pro build 26200 with an AMD Radeon RX 5700 XT. Findings are scoped to that environment and to the three interfaces tested. OpticQuiz is independent and not affiliated with or endorsed by Microsoft. Educational only — not a medical diagnosis.

← More guides

The correction method compared against here is OpticQuiz's published, open-access work — read the paper & open source →