DJI Inspire 3 Zenmuse X9-8K Air Lab Test – Rolling Shutter, Dynamic Range and Latitude

Lab Test

by Gunther MachuJuly 11^th, 2023

It doesn’t happen often, but here is a first: our first drone camera lab test has arrived! When we heard about DJI’s newest addition to the Inspire line back in April with an 8K full frame sensor in combination with ProRes RAW on board – the Zenmuse X9-8K Air – we couldn’t resist putting it through the paces in our lab. Curious to see our results? Then read on in this Inspire 3 Lab Test …

Editor’s note: The DJI Inspire 3 with Zenmuse X9-8K results from this Lab Test have also been added to our Databases. Recording modes and times will be added to the Databases later.

A lot has been written about the DJI Inspire 3 drone with the new Zenmuse X9-8K Air camera module. You can read the announcement with all the specs by my colleague, Jakub Han, here, and another article with everything you need to know here. Also, CineD awarded the NAB 2023 Best of Show Award to DJI – read about it here. A more in-depth video review of the Inspire 3 is currently in the works as well.

Hence, we were quite curious to see how it would fare in our standardized lab test. Now, this was not a straightforward task. We needed to align it properly with our Xyla21 chart (no tripod mount available, obviously ;-), and, it came with the DL 50 mm F2.8 LS ASPH lens. Hence, Florian and I were unsure if our studio light setup would be strong enough to bring the red channel close to clipping at our subject’s face using the F2.8 aperture to test the exposure latitude. Thank you Florian for all your help!

So without further ado, let’s jump straight to the results! For the record, the camera firmware was 10.00.15.01 (most recent at the time of writing).

Rolling shutter of the Inspire 3 Zenmuse X9-8K Air camera module

First, we tested the rolling shutter in full frame mode with the 8K DCI (8192×4320) settings at 25 frames per second (recorded in ProResHQ 422):

Ouch – a whopping 31.3ms (less is better)! That is unfortunately a very bad value. Recent full-frame cameras that we tested showed a range from 3ms to 16ms (e.g. the Sony Venice 2 in 8K DCI has less than 3ms, the Sony a7S III has 8.7ms in 4K, the Canon EOS R3 at 6K has 9.3ms, and the Nikon Z 9 came in at 14.5ms in 8.3K).

You might argue that the onboard gimbal mitigates a lot of the issues with rolling shutter, but as this drone can fly as fast as 95km/h, be aware that rolling shutter effects will show up, especially on tracking shots.

Now, what we found strange is that DJI advertises up to 75fps in 8K for ProRes RAW shooting (in 1:2.4 aspect ratio). How is that possible? We tested it and found that the sensor switches to another readout mode if you go beyond 30 frames per second. From 30 to 60fps, the rolling shutter decreases to 16.3ms. We were a bit unsure about what actually changed, so we took a quick look at our resolution chart to check if the full sensor was still sampled or if some kind of upsampling from a lower resolution was happening:

Regarding resolution, we couldn’t see any difference, even standing a bit further away from the chart to see if the finest details would still be recorded. You can see in the above display that there is no difference in 25 or 60p for 8K DCI ProRes RAW recording.

We cannot tell you what is happening and why the sensor changes readout mode above 30p, but at least there is no impact on the resolution. Hence, for fast action scenes, it is advisable to switch to frame rates higher than 30fps in 8K DCI ProRes RAW mode to mitigate rolling shutter effects.

Interestingly, for ProRes HQ recording mode in full-frame mode, you can also select higher frame rates, but the sensor switches to a different read-out mode AND a reduced 4.1K resolution above 30fps, and again the rolling shutter clocks at 16.3ms.

Now let’s have a look at Super35 (or APS-C) mode:

In 4K DCI in the ProresHQ 422 Super35 settings, the rolling shutter reduces to 14.7ms for 25fps or 50fps.

Dynamic range of the Zenmuse X9-8K Air camera module

We switched to ProRes RAW at 8K DCI, ISO800, and used Apple’s compressor to transcode the ProRes RAW file into 12-bit ProResXQ 4444 D-Log before ingesting frames into IMATEST.

The waveform shows a solid 13 stops above the noise floor:

IMATEST calculates a solid 12.1 stops at a signal-to-noise ratio (SNR) of 2, and 13.4 stops at SNR = 1. Those are very good results, on par if not better than most recent full-frame cameras that we tested in RAW mode. The leader of the pack remains the ALEXA Mini LF, which clocked 13.4 / 14.5 stops at SNR = 2 / 1 in ARRIRAW.

The Zenmuse X9-8K Air camera features a dual native ISO sensor, so we tested the second native ISO at ISO4000 – showing 11.7 / 13.2 stops at SNR = 2 / 1:

Again, there is very little impact on the dynamic range if the second native ISO is utilized – very good!

Now, just out of curiosity, we also tested ProResHQ 422 at ISO800. Here is the waveform and IMATEST result at ISO800 (showing 12 / 13.5 stops at SNR = 2 / 1) – very similar, but a less noisy noise floor:

In ProRes RAW (12-bit) as well as ProResHQ 422 (10-bit), there are additional stops visible in the middle diagram above the blue “13.5” curve (those are stops buried in the noise floor). Potentially, those can be utilized using advanced post-production techniques (noise reduction, etc..).

Unfortunately, the noise present in the noise floor is not evenly distributed, even with the 8K sensor. As a result, at least in our standard studio latitude scene, we were unable to utilize this potential.

Latitude of the Zenmuse X9-8K Air camera module

Latitude is the capability of a camera to retain details and colors when over-or underexposed and pushed back to base exposure. Some time ago, we chose an arbitrary value of around 60% luma (in the waveform) for our subjects’ foreheads in our standard studio scene. This CineD base exposure should help our readers to get a reference point for all the cameras tested, regardless of how they distribute the code values and which LOG mode is used.

For the following shots, we post-processed the ProRes RAW files two different ways and brought them into DaVinci Resolve 18.4 for further analysis (as DVR does not support ProRes RAW so far):

• The Raw Converter App in the Windows App store, which transforms ProRes RAW into a Cinema DNG sequence that can be imported into DVR
• Final Cut Pro, using the ProResRAW dialogue to adjust exposure and exporting a ProResXQ 4444 12-bit file, and then using the official D-Log to Rec709 LUT in DaVinci Resolve

The DNG files from the Raw converter were then developed directly to Rec.709 in the Camera RAW tab of DaVinci Resolve:

Next, we found that our studio light was not able to bring the red channel on my colleague Nino’s face close to clipping at ISO800, due to the rather slow F2.8 lens from DJI. The red channel is exactly 0.5 stops below clipping (as we deducted from comparing it to the waveform plot above):

Thus, we had to add a half stop to the latitude results shown below. Here is a shot at base exposure using the Cinema DNG workflow:

Now, let’s underexpose. We did this by doubling the shutter speed from 1/25s to 1/50, 1/100, and so on until 1/1600. After that, we proceeded to decrease the aperture of the lens step by step, going from F2.8 to F4, F5.6, and finally F8.

Here is 4 stops below the base exposure image, brought back using the CDNG workflow:

We are now at 7.5 stops of latitude (3 over to 4 under plus 0.5). Noise starts to creep into the image, as can be seen in the shadows on the lower right-hand side, for example. Also, shadow areas have become greenish.

This noise can be adequately removed using noise reduction in DaVinci Resolve. Now let’s go to 5 stops of underexposure, brought back (using the CDNG workflow):

The image is now heavily affected by noise, and it is, unfortunately, a rather coarse noise, not finely dispersed, which makes it difficult to remove (see below). Horizontal stripes also start to appear, which noise reduction cannot fully mitigate:

There is a greenish tint to the image which cannot be removed sufficiently. It looks like we have reached the limit. As the skin color on the shadow side of Nino’s face is still somewhat intact, I will include this result.

Let’s look at one stop further to 6 stops of underexposure, brought back using the CDNG workflow:

The image turns greenish (especially in the shadows), and noise degrades the image to a point where it cannot be recovered, even when using heavy temporal and spatial noise reduction (see DVR settings below). Also, the skin color on the shadow side of Nino’s face is not intact anymore:

That yields 8.5 stops of exposure latitude – among the best we have seen from consumer full-frame cameras so far. The ALEXA Mini LF showed 10 stops in comparison (and the ALEXA 35 exhibited 12 stops).

That was the Raw Converter / Cinema DNG / Resolve RAW tab workflow. Now out of curiosity we also tried the second method, developing the ProRes RAW files in Final Cut Pro, doing the exposure adjustment, and then exporting a ProRes XQ4444 12-bit file to DaVinci Resolve.

Well, the official D-Log to Rec709 LUT introduced a more pinkish tint to the image overall, but basically, the results are the same (as expected). Have a look at the 5 stops under, pushed back image below:

Using noise reduction yields the following image:

And just for reference, at 6 stops under the image, noise becomes too coarse to be effectively eliminated with noise reduction. A strong greenish cast covers the image, creeping in from the shadows – plus we have the horizontal lines issue.

Summary of the Inspire 3 Lab Test for the Zenmuse X9 Air

The DJI Inspire 3 Zenmuse X9 Air camera exhibits a rather diverse spectrum of attributes from good to bad in our standardized lab test: the rolling shutter is seriously bad in 8K DCI mode below 30 frames per second but improves significantly above 30 fps. The dynamic range is on par if not better than all full-frame cameras in the consumer price bracket we have tested so far. Only high-end cinema cameras like the RED V-Raptor 8K or the ALEXA Mini LF manage to top these results.

A similar conclusion is derived from the exposure latitude test: a solid 8.5 stops of exposure latitude is shown, which is among the best we have tested in the consumer price bracket.

All in all, a strong performance!

Have you worked with the DJI Inspire 3 and the Zenmuse X9 Air camera so far? Let us know your thoughts and experiences in the comments section below …