Luminance (in)stability in OLED monitors
Topic
It is well known that OLED monitors down-regulate image luminance if, otherwise, the average image luminance would exceed a limit that cannot be handled by the monitor on the long run. However, luminance variation caused by this such limitation shall not be discussed here. Instead, this discussion is about luminance stability when the monitor is not only operated "uniform brightness" mode but also with the brightness set well below the specified maximum for the uniform brightness mode.
One major difference between LCD and OLED technology is that, in LCD monitors, the light comes from a constant backlight and the pixel cells just control how much of this backlight can pass, whereas in OLED monitors, the light is directly generated by the pixels. Assuming that the bulk of the monitor's energy consumption is spent on the pixels' luminous output, the energy demand in LCD monitors is rather constant and much easier to control than in OLED monitors, where energy demand is tightly coupled to the image content and can rapidly change from one image frame to the next. Moreover, this control of the overall OLED current not only has to be fast but also very accurate, given that there is no cancelling out of errors through averaging across the OLED pixels – that is, a 1% error in the overall current directly results in a 1% pixel luminance error for each single pixel. The practical relevance of potentially unstable pixel luminance totally depends on the application, that is how fast and to what extent the (intended) image luminance is changing, and how stable the luminance of a potential target stimulus has to be.
Luminance measurement method
Luminance measurements were taken using a photodiode (OSRAM SFH2240-A01, photo-sensitive area: 2.65x2.65 mm) which was connected to a micro-controller (RaspberryPi RP2350) via a trans-impedance amplifier. All the electronics was USB-bus powered, albeit through linear voltage regulators. The RP2350 offers a 12bit analog-digital converter, which was operated at &tickapprox;385 ksps. However, data was down-sampled to 20 kHz at 16bit before making it available to the PC. The photodiode, which was housed in a small enclosure together with the other electronics, was placed over a 5x5 mm target stimulus at the screen center, at a distance of about 2.5 mm from the screen surface.
A thin foam rubber layer applied to the enclosure's surface, surrounding the enclosure's photodiode cutout, reliably shadowed ambient light. Moreover, the target stimulus was always drawn on a black 35x35 mm square for preventing light emitted by the global background pattern from bleeding into the target area through the mechanical screen structure covering the pixels.
Although the used photodiode is somewhat Vλ-corrected, this correction is far from perfect. Moreover, the setup did not include an optical lens which would have restricted the otherwise relatively wide angular reception profile, which makes the setup more sensitive to different angular emission characteristics of different monitor technologies (e.g., TN vs. OLED). Therefore, the measured signal amplitude for a calibrated D65 white target of 100 cd/m2 still depended on the monitor under test. The signal amplitude for the 5x5 mm target was between 23%FSR and 33%FSR, while being about 30% higher for the bigger 35x35 mm target used in one of the tests.
The stimulus presentation and final luminance sampling followed a 4 Hz clock. In essence, each 4 Hz cycle started with updating the screen image buffer, whether the image content actually changed or not. Photodiode samples that happened to fall in this sampling cycle were averaged, ignoring samples that could potentially be contaminated by the content of the previous or the next image, assuming an image settling time that depended on the monitor technology (e.g., a generous 10 ms for OLED monitors, or 40 ms for the (fast) TN and IPS monitors).
Side note: Originally, an industrial camera (IDS UI-3360CP-NIR, 2048x1088 pixels, 2/3" monochrome CMOS sensor) was used for taking luminance measurements, with the 5x5 mm target stimulus filling an ROI of 1024x1024 camera pixels. However, the final measurement noise with this setup (i.e., in the best-case scenario test with the best monitor) was about one order of magnitude worse than with the photodiode (SD=0.072% vs 0.0054%).
Test stimuli
There were several tests, each with its specific stimulus configuration comprising of a measured target at the screen center and the background covering the rest of the screen. In principle, the measured target was either small or big (5x5 or 35x35 mm), and its luminance could be constant over time or changing. The background either changed between different static patterns (see Figure 1), or it was just all black or all mid-gray throughout, or it was slowly ramped up from black to white and then back down from white to black.
Test protocol
The entire test sequence was run automatically, but the single tests were otherwise independent. Before each test, a quick gamma curve was measured, basically for inferring the pixel values that needed to be programmed for specific luminance levels.
For each of the 6 tests, 5 repetitions of 2 to 4 minute long sweeps were measured. The stimulus sequence within a sweep, even if randomized, was exactly the same for all sweep repetitions (and for all monitors). For example, if the background was changed randomly from time to time, this random sequence of background patterns and the duration of each pattern was kept the same for all sweeps for all monitors. Before each sweep, a black screen was presented for 10 seconds, which was supposed to somewhat reset the state of the screen pixels and monitor control electronics. For a single monitor, the data collection took about 95 minutes.
Analysis
Very slow luminance changes are not of interest here, which is why luminance drift was removed from each sweep. The drift was modeled by a smoothing quadratic spline with support points about every 60 s. For 2-minute sweeps, for example, this resulted in 4 coefficients per sweep, which included the coefficients for the spline boundary conditions at the sweep start and end. Because such drift model is too flexible for preserving systematic luminance changes which are of interest in the BgRamp+TgtMaxSmall test (see below), the spline-based drift model was replaced by a simple linear drift model for this particular test.
The blue vertical line indicates where the luminance reaches 50% of the maximum luminance. All tests use target luminances higher than 50%, where the relative luminance step sizes are about 1% and are essentially independent of the monitor's white contrast.
Also of no interest are the absolute luminance levels, which is why the luminance error is expressed as relative luminance error, i.e., the difference between the measured and the expected luminance, divided by the expected luminance. The expected luminance was defined as the average luminance across all measurements taken within a sweep for the respective target luminance. This definition of relative luminance error closely reflects what is also perceptually relevant; a 1% relative luminance error is perceptually as large in dark stimuli as in bright stimuli, even though the corresponding absolute luminance errors differ.
For the quantitative evaluation of the observed relative luminance error values, a comparison with the relative luminance step size following a gamma transfer function might be useful. Unfortunately, the step sizes depend on the assumed gamma value, white-contrast, and specific pixel value. Figure 2 shows curves for gamma=2.2 and three different contrast levels. For the test described here, which mostly use a 100%-white target or at least a target with a luminance above 50% of maximum white, an 8 bit step corresponds to a relative step size of roundabout 1%. A uniform distribution of an according luminance error would have an SD of 1%/sqrt(12) ≈ 0.3%. This means that, at least for higher pixel values, the relative luminance error resulting just from the 8 bit quantization is 0.3%. However, this a very theoretical value and does not include the additional round-off noise introduced by the color processing in the monitor.
Results overview
| Bg0 + TgtMaxSmall |
BgMid + TgtMaxSmall |
BgVar + TgtMaxSmall |
BgRamp + TgtMaxSmall |
Bg0 + TgtVarSmall |
Bg0 + TgtVarBig | |
|---|---|---|---|---|---|---|
| BenQ XL2540 (TN) |
0.028% (±0.0020) |
0.015% (±0.0021) |
0.085% (±0.00077) |
0.056% (±0.0011) |
0.027% (±0.0030) |
0.023% (±0.0024) |
| ASUS XG27AQ (IPS) |
0.014% (±0.00055) |
0.015% (±0.00052) |
0.050% (±0.00026) |
0.034% (±0.0032) |
0.015% (±0.00044) |
0.015% (±0.00046) |
| Razer Raptor 27 165Hz (IPS) |
0.0054% (±0.00020) |
0.0059% (±0.00028) |
0.041% (±0.000068) |
0.019% (±0.00064) |
0.011% (±0.00019) |
0.011% (±0.00029) |
| ASUS PG27AQDP (W-OLED) |
0.017% (±0.0012) |
0.018% (±0.00072) |
0.17% (±0.00093) |
0.11% (±0.0047) |
0.022% (±0.00045) |
0.024% (±0.00093) |
| ASUS PG27UCDM (QD-OLED) |
0.075% (±0.0095) |
0.046% (±0.0029) |
2.2% (±0.013) |
2.7% (±0.013) |
0.079% (±0.0055) |
0.10% (±0.016) |
| MSI 271QRX (QD-OLED) |
0.066% (±0.0085) |
0.037% (±0.0084) |
0.31% (±0.0060) |
0.79% (±0.033) |
0.066% (±0.0076) |
0.066% (±0.0076) |
Table 1 and Figure 3 show the relative luminance errors in terms of standard deviations (SDs) for the test conditions explained further below. Note that for the MSI 271QRX monitor (violet bars) the small target was actually always replaced by the big target, which – the big target – was otherwise only used in the last test condition (Bg0+TgtVarBig). This was necessary because the pixel shift (OLED care option) could not be disabled for this monitor and prevented measuring the small target with sufficient consistency.
The TN and an IPS monitors were mainly included for reference, whereas the main focus lies on the comparison of the three OLED monitors, specifically the difference between W-OLED and QD-OLED.
Notice that the Y axis in Figure 3 is log-scaled, so the performance differences are actually bigger than they appear in the figure.
Test scenarios
Best-case (baseline)
Test: Bg0+TgtMaxSmall and BgMid+TgtMaxSmall.
Obviously, the screen image should remain completely unchanged for this test. One cannot be sure though whether a black background is really optimal here, because with a basically all black screen, the monitor's control circuit is operating at the lower boundary of its operating range, where the circuit might be working less optimal than further away from the boundaries. Therefore, a mid-gray background might be more favorable.
The Razer monitor shows, with relErr=0.0054% in the BgMid+TgtMaxSmall condition, the best result, obviously also providing an upper limit for the photodiode measurement noise, at least for this luminance level, i.e., at 100 cd/m2. Further measurements not shown here suggest that this value might reflect, at least to some relevant extent, photodiode measurement noise rather than monitor luminance noise. Anyway, given that all other results are much higher than relErr=0.005%, those results are not affected by measurement noise.
The very different performance of the two IPS monitors (Razer with 0.0054% vs. ASUS with 0.014% – almost a factor of 3) shows how much the particular implementation can play a role. In this case, even the panels come from different manufacturers (Innolux vs. AUO).
It is noteworthy that the major part of the noise exhibited by the worst-performing monitors (i.e., the TN and both QD-OLED monitors) originates from low-frequency step-wise changes, especially for the two QD-OLED monitors (see Figure 4).
Effect of background pattern
Test: BgVar+TgtMaxSmall.
Whereas the previous test was about the best-case scenario, this test is more about the worst-case scenario. How is the target stimulus luminance modulated by the background pattern and average luminance? The effect of the background on the target luminance might not only be a result of different average luminances but also where on the screen, with respect of the target position, either dark or bright regions are presented. There are many potential factors at play, but the goal is not to identify and isolate them but to create a worst-case scenario, which gives each of these factors a fair chance to come into effect. This is done by presenting different patterns (see Figure 1) in random order and for random duration, all the while the target stimulus at the screen center remains the same. Importantly, this includes scenarios where the average luminance will change a lot within a short time, allowing for potential settling effects of the monitor's control circuit to surface. In fact, the pattern sequence includes more full black and full white screens than other patterns for making more extreme changes in the average luminance more likely.
Figure 5 (and Figure 6 for a zoomed-in version) shows the traces for the OLED monitors.
Notice the different Y axes scalings, which differ by an order of magnitude. This is how far off the ASUS QD-OLED is in this test compared to the other two OLED monitors. Also notice that the stimulus time course for the background (blue curve in the first panel) indicates the background pattern and not the background pattern's luminance (at least not directly). For the pattern indices, see Figure 1. Therefore, there is no obvious low/high correspondence expected between the blue curve and the traces. For judging the magnitude of the Y values, be reminded that an 8 bit step corresponds to a relative step size of roundabout 1%, meaning that, for the ASUS QD-OLED monitor, the dominant modulation amplitude is in the order of several 8 bit steps – not great!
There is little low/high correspondence between the traces of the different monitors to begin with, but the ASUS QD-OLED in particular exhibits a pronounced settling behavior whenever the relative luminance error (and, thus, the luminance) changes from a higher to a lower level, which becomes clearer when zooming in (see Figure 6). That is, the settling behavior is even uni-directional. Whatever the explanation might be, this is not specific to QD-OLED technology, because the MSI QD-OLED monitor does not show this behavior at all.
Figure 7) shows that the patterns of non-OLED monitors look rather similar to those of the OLED monitors, albeit the error levels are mostly lower. That the patterns look similar does not necessarily mean that they are caused by the same mechanism. At least for the non-OLED monitors, it seems very unlikely that the observed luminance modulation is reflecting a modulation of the LED backlight, which operates very independently of the image content. Although this is clearly different in OLED monitors, as explained in the introduction, it is just an assumption that the luminance modulation observed for OLED monitors is dominantly caused by the overall OLED current control circuit which, in turn, is affected by the average image luminance. To isolate average luminance effects from background pattern effects, the next test is better suited.
Effect of average luminance
Test: BgRamp+TgtMaxSmall.
This test probes the monitor's behavior for different image average luminances without potentially contaminating the results by additional background pattern effects as is the case with the previous BgVar+TgtMaxSmall test. Moreover, this test covers the range of average luminance exhaustively, which can reveal different operating regimes of the monitor's control circuit. To do so, the entire background (excluding the target region) is slowly changed from black to full white and back. Having both, the up-ramp and the down-ramp, allows potentially interesting symmetry effects to show. Note that the measured target is a full white stimulus, which might be considered the best-case scenario as far as relative luminance errors go.
br/>
Figure 8 shows the traces for the OLED monitors. The W-OLED monitor stands out positively, as it shows a smooth modulation similar to the non-OLED monitors albeit at a somewhat higher error level (see Figure 9). The ASUS QD-OLED is the worst, although mostly because of the behavior at average luminances below ≈40% (which corresponds to 40 cd/m2 under the test conditions used here). The behavior of the MSI monitor seems not as reproducible as for the other monitors, neither absolutely nor in terns of symmetry between the up- and the down-ramp. This might indicate a high susceptibility to temperature fluctuations. However, this effect is not dominating the behavior, and the errors are still smaller than for the ASUS QD-OLED monitor.
Also the non-OLED monitors show a modulation with the background luminance (Figure 9). However, this modulation might originate from more local interactions between target and background regions rather than from the change of overall average luminance. If such interaction exists, the associated luminance errors are likely modulated by the local background luminance levels rather than some overall background luminance. This hypothesis is supported by comparing error levels between different background patterns while taking also the average luminance of the respective average luminances into account. The simple error vs. average luminance relationship non-OLED monitors show in the ramp test makes such dissociation between average luminance effects and background pattern effects relatively easy, in contrast to the OLED monitors (at least the QD-OLED monitors), where both tests, the ramp test and the background pattern test, show complex patterns. Be it as it may, the exact cause of the errors is of less interest here than the error magnitudes and patterns.
Speaking of patterns, the TN monitor stands out because of the many spikes in the curves. Moreover, the curves are upside-down when compared to the other monitors. The sign of the errors might reflect that TN panels are usually "normally black" (i.e., the pixel get brighter as pixel voltage increases), whereas IPS panels are "normally white".
Target luminance repeatability
Test: Bg0+TgtVarSmall.
The tests described so far were designed to probe the modulation effect of the background on a constant target stimulus, whereas this test is looking at the luminance stability or repeatability within a sequence of different target luminances. Since the small target was used for this test, the overall average luminance was considered basically constant, thereby going easy on the power control circuit. This test aims at revealing potential memory effects of the pixel driver electronics.
The stimulus is similar to what is normally used for settling time and overshoot measurements, where the measured target's luminance is switched between predefined luminance levels. But those tests are measuring short-term settling effects and use the settled luminance state after each luminance switch as a reference. Here, in contrast, it is about how stable or reproducible the settled states are in different temporal contexts. The this end, the target stimulus is changed between different luminance levels and for random time intervals. For the sake of high SNR, only luminance levels above 50% are probed.
A black background is used here because it is assumed to be the worst-case scenario, being well aware that this assumption might be wrong. However, the first test of the test ensemble (i.e., Bg0+TgtMaxSmall vs. BgMid+TgtMaxSmall, see above) and the ramp test (i.e., 'BgRamp+TgtMaxSmall) provide some evidence that a black background is at least not beneficial when it comes to luminance errors, at least not if the target is full white as used in the mentioned tests.
The results are presented in the same form as above, i.e., as relative errors over time, even though this is somewhat misleading. This is because abrupt changes in relative errors as well as the high variance of relative errors across measurement sweeps are merely artifacts of the presentation from. The different target luminance levels require accordingly different reference luminances for computing relative errors. Therefore, not only time is changing when going from one target luminance to the next but also the reference value used for computing the respective relative errors, thereby introducing potential jumps in the traces. Moreover, small differences between the sweeps can translate to apparently less repeatability between measurement sweeps. Another aspect concerns the error magnitudes. Just be reminded that, if the absolute errors happen to not scale with the luminance level, lower luminance levels will come with higher relative errors just because lower luminance levels mean lower reference values the errors are related to. This is an intended effect, because relative errors are just the better measure when it comes to the perceptual impact of, as explained in the Analysis section. All these issues make the interpretation of the graphs difficult.
That being said, Figure 10 and Figure 11 show the results for the OLED and non-OLED monitors respectively. The pattern exhibited by the QD-LED monitors set them apart from the rest of the other monitors by showing much more non-random components. Due to the issues just explained, it is hard to extract anything more from this data than the total noise levels. And the noise levels are actually not too far from the levels observed for the nest-case scenario (Bg0+TgtMaxSmall, see Figure 3), which is a good thing. The difference between these tests could actually be explained just by the relative errors inversely scaling with the target luminance (not saying that this is necessarily the case, but it could be). The only exception is the Razer monitor (IPS), which shows a significant lower error in the best-case scenario, even though it still shows the lowest error of all the tested monitors also in this less favorable scenario.
Effect of local pixel neighborhood
Test: Bg0+TgtVarBig.
This is the same test as before but with the bigger target, for testing whether this would stabilize the central area of the target. With the used photodiode, the measured area was not exactly confined to the area covered by the small target, but the measured area was far smaller than the big target, thereby providing plenty of pixels around the measured area for having a stabilization effect on the central area without contaminating the measurement. If there was such a stabilization effect, errors should be smaller than for the previous test (Bg0+TgtVarSmall). On the other hand, a bigger target has a higher impact on the total average luminance, which could have a potentially negative impact. This is what might be happening with the ASUS QD-OLED monitor (see Figure 12), whereas there is no significant change in errors for the other monitors between small and big target (see Figure 3).
Conclusion
Given the small number of monitors tested, the findings presented here cannot be generalized. Also, pitting non-OLED monitors against OLED monitors seems unfair, because at least the key tests used here (variable background and ramping background) were designed to tap into a problem mainly OLED monitors have: power control. Of course, thing would look differently when testing for pixel-inversion or IPS glow. Nevertheless, the non-OLED data seemed to be interesting enough to include them here, and if it is only to show how much variability there is even across seemingly similar monitors like the two IPS monitors.
When just looking at the OLED monitors, one could think that W-OLED is just the superior technology when it comes to luminance stability. But the sample size is just too small, and the performance difference between the two QD-OLED monitors show how much implementation matters, let alone that it would be hard to explain why W-OLED should be better than QD-OLED in this regard. Even claiming that OLED monitors in general have a luminance stability problem is more supported by the potentially good explanation (power control) than by the data presented here (again, because of the small sample size).
It is difficult to evaluate the practical relevance of the presented findings, even when assuming that luminance stability would be the only criterion that matters. For example, if predictability of luminance stability would be of most importance, then the TN monitor should be rated best, because it is performing at the same mediocre level throughout – very noisy even in the best-case scenario, but not so much affected by the image content either. On the other hand, the Razer IPS monitor should be rated worst, even though it performs better than all the other monitors in each single test. As so often, the final verdict depends on the application.