Finally, we were able to test the first proper video-capable, medium-format camera in our lab, the FUJIFILM GFX100 II. Using the new F-Log2 image profile and robust internal codecs (4:2:2 10bit ProRes HQ), in theory, all the preconditions should be met for fantastic results. Or not? Curious to hear how it turned out? Then read on …
The FUJIFILM GFX100 II is a really unique offering, especially in terms of sensor size and resolution, and this camera very justifiably won CineD’s “Camera of the Year 2023” title, together with the iPhone 15 Pro / Max (read our article here). Also, my dear colleague Johnnie made one of his lovely mini-docs with it, and you can read his review here.
My colleague Florian (again, thank you for your help) and I were curious to put the GFX100 II through the standard CineD lab torture test.
The GFX100 II offers a huge variety of sensor read-out modes for video. The FUJIFILM site has more information, but for your convenience, here are the GF options in medium format. There are Premista, 35mm (full frame), and anamorphic options as well.
As you can see, the only mode that covers the full sensor width in 16:9 / 17:9 is the 4K mode. The other modes are either horizontally or vertically cropped. We conducted a range of tests to assess rolling shutter and dynamic range across different modes, utilizing our Xyla21 chart. The corresponding graphs can be found in the CineD database.
To me, any mode that does not use the full sensor width defeats the purpose of testing a medium-format camera. Hence in this lab test, we focused on the GF 4K mode (Cine 5.8k uses the 2.35:1 aspect ratio, which does not comply with our 16:9 standard studio scene latitude test requirement).
Plus, there is the “F-Log2 D RANGE PRIORITY” “ON” or “OFF” feature. FUJIFILM has not specified what it does exactly, but I speculate that in D RANGE “ON” the full sensor resolution (11648 x 7768) is downsampled to the respective video resolution, at the cost of a higher rolling shutter. Hence, in principle, the 4K (3840×2160) GF F-Log2 D RANGE PRIORITY ON settings should yield the lowest noise, hence the best dynamic range results, utilizing the full sensor width in a standard 16:9 ratio.
As for the lens, we were not able to use our standard Zeiss Compact Prime 85mm T1.5 lens, so we rented the FUJINON GF 80mm f1.7 R WR instead. A superb lens, sharp even when wide open.
Rolling shutter of the FUJIFILM GFX100 II
Fine – let’s get going! The first test with our 300Hz strobe light is the full sensor width, full downsampled 4K mode at 25p, D Range ON. We get a result of 26.5ms (less is better):
We have no comparison with other medium format sensors, but for video work, this is on the (very) high side of things.
Let’s have a look at the same mode, but D RANGE PRIORITY OFF:
As can be seen above, the full sensor resolution downsampling takes its toll on the rolling shutter. In D RANGE PRIORITY OFF, it seems that some sort of line skipping is happening, allowing a much faster sensor readout – potentially at the expense of dynamic range and exposure latitude. We will dive into that in the next section.
To give you a reference, recent consumer full-frame cameras like the Sony A1, Canon EOS R5 C, or Nikon Z 9 are all around 15ms read-out speed. The Canon EOS R3 clocks 9.5ms, and the king of consumer full-frame cameras, the Sony A7S III, only has 8.7ms. (Excluding the Sony A9 III with its Global Shutter image sensor).
In full sensor width GF 5.8K mode, we get a rolling shutter of 25.9ms. I would have expected a faster readout as the picture height is less (2.35:1). One more result: in cropped GF 8K mode the rolling shutter is 31.7ms.
Dynamic range of the FUJIFILM GFX100 II
As usual, for the dynamic range and latitude tests, we used the camera settings that allows minimum noise reduction as it can not be turned completely off.
Let’s have a look at the full sensor width 4K D RANGE PRIORITY ON waveform first. If you are not familiar with how we test dynamic range, have a look here.
The 4K F-Log2 D Range ON waveform using 10bit 4:2:2 ProRes HQ at the native ISO800 yields a solid 13 stops above the noise floor. The noise floor looks super clean, probably as a result of the downsampling of the massive native sensor resolution (11648 x 7768) to 4K:
The corresponding IMATEST result yields high values of dynamic range: 12.4 stops at a signal-to-noise ratio (SNR) of 2 and 13.7 at SNR = 1:
Now, what happens if we turn D RANGE PRIORITY to OFF? The waveform looks a bit noisier, and IMATEST yields 11.7 stops at SNR = 2 and 13 stops at SNR = 1. That’s 0.7 stops worse:
If we switch to the (cropped) 8k mode, probably the closest to native sensor resolution, IMATEST calculates 11.3 stops at SNR = 2 and 12.8 stops at SNR = 1. This can be regarded as the “native” dynamic range of the sensor, without any effects of downsampling.
Head over to the CineD camera database for additional IMATEST results for the 5.4K Premista (12.8 / 13.9 stops at SNR = 2 / 1) and 4.8K 35mm modes (12.4 / 13.9 stops).
Exposure latitude of the FUJIFILM GFX100 II
As stated before, latitude is the capability of a camera to retain details and colors when over- or underexposed and pushed back to a base exposure. This test is very revealing, as it pushes the complete image pipeline of any camera to its absolute limits – not just in the highlights but mostly in the shadows.
Our studio base exposure is (arbitrarily) chosen as having an (ungraded) luma value of 60% on the forehead of our subject on the waveform monitor – in this case, my colleague Johnnie. We developed the ProResHQ files using the official GFX100II_FLog2_FGamut_to_WDR_BT.709_33grid_V.1.00 LUT, which is available on the FUJIFILM homepage:
Again, we used 4K F-Log2 D RANGE PRIORITY ON mode using 10-bit 4:2:2 ProResHQ mode.
Now let’s see where the red channel starts to clip on Johnnie’s forehead. It begins at 3 stops above the base exposure, as can be seen below. Looking at the RGB waveform of the ungraded clip below, the red channel on Johnnie’s forehead appears intact. However, in the developed version and the image pushed back 3 stops, it starts to look a bit overexposed.
This is something to consider when exposing the GFX100 II, as F-Log2 has a rather smooth transition towards highlights, which in turn makes it a bit more difficult to assess the exact clipping point:
Now, let’s see how far we can underexpose and bring back the image. Not much is happening between base exposure and 4 stops under, although at 3 stops under (6 stops of exposure latitude), sensor smear becomes apparent. It can be seen more easily at 4 stops under, so let’s jump there:
Now noise starts to kick in. We are at 7 stops of exposure latitude. In the shadows, the image starts to degrade, but the shadow side of Johnnie’s face (my most important criterion) still looks OK, even without applying further noise reduction.
However, something that to my eyes looks like sensor smear becomes visible. The white piece of paper on the left leads to a distinctive horizontal band across the image all the way to the right where the image turns greenish, everything else pinkish (within the two horizontal boundaries of the white piece of paper, clearly visible by the two horizontal lines). Not good. However, luma noise is very finely distributed.
Let’s move to 5 stops under, pushed back to base:
More noise is appearing, but the shadow side of Johnnie’s face is still mostly intact, even without noise reduction. Overall, this would be a good result at 8 stops latitude. However, sensor smear is even more pronounced, and in the moving image, large blotches of pinkish chroma noise can be seen. Signs of color banding are also appearing. Have a look at the shadow on the left side in the background of the color checkr.
Noise reduction helps here and cleans up the image nicely – however, the sensor smear is even more apparent and is happening across the image (visible as horizontal lines):
Now, 6 stops under:
Noise reduction cannot save this image:
Now, to stay consistent with earlier lab test assessments that are already at 4 stops under (7 stops of exposure latitude), the sensor smear is impacting the image in a way that cannot be fully recovered, although noise in the image is still well-controlled. That gives a total of 6 stops of exposure latitude.
Now, sensor smear can happen if the pixel density on a sensor is very high and an electrical charge impacts neighboring pixels (in this case due to the harsh horizontal transition from the grey background to the white piece of paper – thus the horizontal lines).
We have seen a similar phenomenon on the Blackmagic URSA Mini Pro 12K camera (lab test here – also see the comments from Alister Chapman and John Brawley) – obviously, sensor smear is more difficult to control the higher the pixel density.
Now, to prove the point we also tested the D RANGE PRIORITY “OFF” settings. In D RANGE “OFF”, the sensor seems to skip lines, hence smear should not affect neighboring pixels.
Let’s have a look at 4 stops under, brought back to base with D RANGE “OFF”:
Much noisier, but no visible sign of sensor smear. The image can be mostly repaired by using noise reduction, however, towards the shadows nasty color banding (stepwise lume / chroma transitions) is starting to appear – hence, even in this mode, 7 stops are the limit. Let’s look at 5 under:
Clearly, much noisier than the D RANGE ON image at 5 stops under (have a look above), but no sensor smear. Very visible are the color banding structures on the left side – the shadow of the color checkr and the area behind Johnnie.
Let’s have a look if this can be saved by noise reduction:
Overall as the noise is still very finely distributed, the image cleans up nicely. Unfortunately, in the shadows, the image shows massive color banding and cannot be reconstructed.
In summary, it can be seen how D RANGE ON leads to a much cleaner, much more usable image if you pull shadows (at the cost of a higher rolling shutter). Unfortunately, this mode is hampered by sensor smear, leading to a usable 6 stops of latitude only. To eliminate the sensor smear, the 4K line skipping mode can be enabled giving a better rolling shutter but much more noise and stepwise transitions into shadows (banding).
If the sensor didn’t show smear, 8 stops of exposure latitude would be easily possible, as the noise is distributed very finely and can be cleaned up in post. To put this into perspective, 8-stop latitude is now the almost standard result for full-frame consumer cameras. However, our benchmark cameras ARRI ALEXA Mini LF display what is possible by showing 10 stops, and the ALEXA 35 has 12 stops of exposure latitude.
Summary
The FUJIFILM GFX100 II exhibits very mixed results in the lab: rolling shutter values are in general very high – as expected for a 102-megapixel medium format sensor. The line-skipping modes (D RANGE OFF) are good in terms of rolling shutter (15ms in 4K) but hampered in the dynamic range and latitude department.
Switching to full resolution downscaling (F-Log 2 D RANGE PRIORITY ON), rolling shutter values are on the high side, but the dynamic range improves. Also, transitions into highlights are very smooth. Nevertheless, in these modes, sensor smear impacts shadow areas. Personally, I would still only use the downscaled modes as the images are simply better.
Hence, don’t expect wonders in the dynamic range department just because of the massive medium format sensor. The GFX100 II is the first solid offering from a manufacturer that enables you to use this camera for serious video work with the “look” of medium format. For comparison, FUJIFILM’s own APS-C X-H2S camera fares much better in the lab test and exhibited close to 9 stops of exposure latitude with a rolling shutter of 9.7ms.
Have you shot with the FUJIFILM GFX100 II yet? What are your experiences? Let us know in the comments below.
