Past 12 Megapixels, Your Eye and Physics Stop Counting

Every new camera launch sells a bigger megapixel number. I want to know where that number stops helping a person who looks at a print or a screen. My answer has two parts. Diffraction caps the detail a lens can deliver at a given f-number. The human eye caps the detail a viewer can use. Both caps land near 12 MP for common cases. I derived the numbers by hand, without the Lab, and I list every input so you can repeat them.

Question

At what pixel count does adding pixels stop adding visible detail to a typical print or screen? And which factor then sets image quality instead?

I hold a prior of 0.7 that sensor size and lens quality matter more than pixel count above about 12 MP. This post tests that prior. It could fail. If the derived caps came out near 40 MP, I would drop the position.

Data and where it came from

I use five sources. I did not run any camera.

• Imatest (a software and lab company) publishes the diffraction formulas I use. The cutoff frequency is $f_c = 1/(\lambda N)$, with wavelength $\lambda$ and f-number $N$. The Airy disk radius is $1.22\,\lambda N$. The diffraction MTF is $(2/\pi)[\arccos(s) - s\sqrt{1-s^2}]$ with $s = f\lambda N$. Imatest also defines $Q = \lambda N / p$ for pixel size $p$, and says that at Q = 2 the cutoff equals the Nyquist frequency [1]. This is measured and standard optics, not opinion.
• A vision study on arXiv measured the resolution limit of the eye. It reports 94 pixels per degree for central grayscale vision, 89 for red-green patterns and 53 for yellow-violet. It says this is above the older 60 pixels per degree standard from 20/20 vision [2]. This is a measurement from human subjects. I have not seen a replication.
• DXOMARK scores sensors after normalizing every image to 8 MP, printed at 8 x 12 inches and 300 dpi [3]. This is a lab choice about what a typical print is.
• DXOMARK's review of the Sony RX100 V reports a 1-inch sensor with 20.1 MP. It says the camera delivers "essentially one stop lower image quality in all categories at base ISO" than the APS-C Sony A6300, and is much closer to the Micro Four Thirds Olympus PEN-F [4]. This is measured.
• DPReview explains that a larger sensor gathers more total light for the same output brightness. So every tone from the larger sensor has a better signal-to-noise ratio [5].

Gap: I did not pull lens MTF charts from a manufacturer or a lab such as Imatest or DXOMARK's lens tests. I derive an ideal lens only. A real lens does worse, never better.

Method

I assume green light at $\lambda = 0.55\ \mu m$. I use nominal sensor sizes: full frame 36 x 24 mm, APS-C 23.5 x 15.6 mm, Micro Four Thirds 17.3 x 13 mm, 1-inch 13.2 x 8.8 mm. These are inputs I assumed from standard format sizes, not numbers from the sources above.

Step 1. Airy disk. Diameter is $2.44\,\lambda N = 1.34\,N\ \mu m$. At f/8 it is 10.7 $\mu m$. At f/16 it is 21.5 $\mu m$. Compare that with pixel pitch. A 24 MP full frame sensor has 6.0 $\mu m$ pixels. A 20 MP 1-inch sensor has 2.4 $\mu m$ pixels. At f/8 the Airy disk covers about 1.8 pixels on the first and 4.5 pixels on the second.

Step 2. Contrast at the pixel limit. The Airy diameter alone is a crude test. I use the diffraction MTF at the Nyquist frequency instead. Nyquist is $1/(2p)$, so $s = \lambda N/(2p)$. I put that into the Imatest formula.

Step 3. Diffraction-limited pixel count. Detail at 50% contrast (MTF50) tracks perceived sharpness. The diffraction MTF falls to 0.5 at $s \approx 0.40$. So $f_{50} \approx 0.40/(\lambda N)$. At f/8 that is 91 cycles per mm. Across a sensor of width $W$ there are $f_{50} W$ cycles, so $2 f_{50} W$ pixels across the width. I then scale to the sensor aspect ratio. This count is the pixel number at which an ideal lens at that f-number still gives 50% contrast at the pixel limit.

Step 4. Eye limit. I take an 8 x 12 inch print (14.4 inch, 366 mm diagonal) viewed at its diagonal. One pixel per $1/94$ degree is 0.068 mm at 366 mm, which is 374 pixels per inch. That gives 4490 x 2990 pixels, or 13.4 MP. With the older 60 pixels per degree, it is 238 pixels per inch and 5.4 MP.

Result with numbers and uncertainty

Contrast at Nyquist, diffraction only (ideal lens)

Sensor	MP	Pixel pitch	f/5.6	f/8	f/11	f/16
Full frame	12	8.5 $\mu m$	n/a	0.67	n/a	0.37
Full frame	24	6.0 $\mu m$	0.68	0.54	0.39	0.16
Full frame	45	4.4 $\mu m$	n/a	0.39	0.20	about 0
APS-C	24	3.9 $\mu m$	0.51	0.32	n/a	n/a
Micro Four Thirds	20	3.3 $\mu m$	n/a	0.23	n/a	n/a
1-inch	20	2.4 $\mu m$	0.25	0.03	n/a	n/a

At f/8 the 1-inch sensor receives almost no contrast at its own pixel limit. The 20 MP count exists on the spec sheet, not in the image. Imatest says system MTF at Nyquist above about 0.3 raises aliasing concerns [search summary, not a quote I rely on], so I treat 0.1 to 0.3 as the useful band for a modern sensor. I mark that threshold as taste, not measurement.

Diffraction-limited pixel count at f/8 (MTF50)

Sensor	Width x height	Pixels at f/8
Full frame	36 x 24 mm	about 29 MP
APS-C	23.5 x 15.6 mm	about 12 MP
Micro Four Thirds	17.3 x 13 mm	about 7.5 MP
1-inch	13.2 x 8.8 mm	about 3.9 MP

At f/16 every figure falls to a quarter. Full frame drops to about 7 MP. At f/5.6 full frame rises to about 59 MP, but only if the lens has no aberrations at f/5.6. Imatest notes that aberrations get worse at large apertures and lenses tend to be sharpest 2 to 3 stops below maximum [1]. So f/5.6 is where an ideal lens helps and a real lens often does not.

The pixel count at a fixed f-number scales with sensor area. That is the result I care about. The sensor, not the pixel grid, sets how many pixels the physics will fill.

Eye cap

The eye cap for an 8 x 12 inch print at diagonal distance is 5.4 MP (60 pixels per degree) to 13.4 MP (94 pixels per degree). The 8 MP print that DXOMARK uses sits inside this band [3]. A 4K screen has 3840 x 2160 pixels, which is 8.3 MP (arithmetic), also inside the band.

Uncertainty

I give a range, not a point. The two caps overlap at roughly 5 to 13 MP for a normal viewing distance. The diffraction cap at f/8 spans 4 MP (1-inch) to 29 MP (full frame). So "12 MP" is a good rule for APS-C at f/8 and for an average viewer, and a loose rule elsewhere. I put the eye number at about 25 percent uncertainty because the 94 pixels per degree result comes from one study [2] and the older 60 comes from a convention.

Where sensor size shows in measured data

DXOMARK measured the 1-inch RX100 V about one stop below the APS-C A6300 [4]. My own light arithmetic gives a different quantity. At the same f-number and exposure, APS-C (366.6 mm²) gathers 3.2 times the light of 1-inch (116.2 mm²), which is 1.7 stops of light. Shot-noise SNR rises with the square root of light, so the SNR gain is about 0.8 stop. The DXOMARK figure is of the same order. I do not claim they match. Read noise, sensor design and lens differences also enter, and DPReview's page does not cover read noise [5]. DXOMARK also describes SNR on a decibel scale where 6 dB is a doubling [3]. I did not check the A6300 against these formulas.

Sensitivity: which assumption moves the result most

I rank the assumptions by how much each changes the answer.

1. Viewing distance. This is the largest. If you view the same print at half the diagonal, pixel need rises fourfold: 54 MP at 94 pixels per degree. A viewer who leans in to look at a 20 x 30 inch print gets a very different answer. Eye caps assume the viewer stands back. Argued.
2. f-number. Diffraction count scales with $1/N^2$. Going from f/8 to f/16 cuts it to a quarter. A landscape shooter at f/16 gets about 7 MP of real detail on full frame. A portrait shooter at f/2 is limited by aberrations instead.
3. Sensor width. The count scales with $W^2$. Moving from APS-C to full frame lifts the f/8 cap from 12 to 29 MP. This is the sensor-size effect, and it is why the same pixel count means different things on different sensors.
4. The MTF threshold. I used MTF50 for the cap. If you accept MTF at 20%, the diffraction count rises. If you demand 50% at Nyquist, it falls. This choice is taste. The measured parts are the formulas.
5. Wavelength. Using 0.45 $\mu m$ blue or 0.65 $\mu m$ red shifts the count by about 20 to 30 percent. Smaller than the others.
6. Sharpening, demosaicing and lens aberration. Software can lift contrast above the optical limit. It cannot create detail the lens did not form. A real lens lowers every number. I did not model these.

What I now think

Measured: diffraction sets a hard ceiling that depends on f-number and sensor size. Measured: the eye resolves between 53 and 94 pixels per degree depending on the pattern. Argued: for a print or screen viewed at normal distance, the useful range is about 5 to 13 MP. Taste: whether the last few percent of contrast at Nyquist is worth 24 MP files.

I do not say more megapixels hurt. A 45 MP full frame sensor at f/8 still carries 0.39 contrast at its pixel limit in my table, and it lets you crop. My claim is narrower. The megapixel number tells you the grid. It does not tell you what the lens and the sensor area fill into the grid. A 20 MP 1-inch sensor at f/8 leaves almost no contrast at its pixel limit.

My prior of 0.7 holds, and I move it to about 0.72. The derivation supports it, but my lens data is ideal and my eye data comes from one study. Three things would change my mind. Published lab MTF curves showing real lenses beating the diffraction ceiling at f/8 on 1-inch sensors would do it. A replicated result putting the eye limit above 120 pixels per degree would do it. A viewing study showing most people look at prints from under half the diagonal would also do it.

One question for you: find a photograph you love, one that you have seen in print. How close can you stand before the detail stops improving? Measure that distance against the print diagonal. That ratio is the one number that decides whether your next camera needs more pixels.