Fujifilm GFX with 63/2.8 AF accuracy

This is the 32nd in a series of posts on the Fujifilm GFX-50S. The series starts here.

There have been conflicting reports about whether the Fujifilm GFX 50S focuses at taking aperture or not. It’s an important question, because, as we’ve seen in the last post, the 63 mm f/2.8 lens has a moderate amount of focus shift, and should be focused at the taking aperture in situations where image sharpness at the focus point is critical. It sure looks like it’s stopping down to focus when you peer into the lens while firing off frames, but it all happens so fast that you may just be looking at the lens stopping down after it achieves focus. 

I’m not going to bury the lede, here. With the 63 mm f/2.8 lens attached, the GFX does not achieve critical focus anywhere near constantly at f/4 and f/5.6, and even f/8 is a little dodgy.

I’ll show you the chart with the smoking gun first, explain it a bit, and then go into painful detail about how it was generated for those of you who either care about such things in the abstract, or need to know what I did to replicate it or find out if it passes your personal muster. 

Here’s the chart:

What you are looking at is a standard measure of sharpness, MTF50, as measured in cycles per picture height (cy/ph) at the whole stops from f/2.8 through f/16. The orange line is the average, or mean, of the 41 exposures at each aperture. The gray line is that average, plus the standard deviation for the 41 exposures, and the blue line is the average minus that standard deviation. If the data were Gaussian (which it isn’t) two thirds of the values would be between the gray and the blue lines. 

A word about the processing. The images were developed in Lightroom with default settings except for white balance, which was set to the target white with the eye dropper. Since the Lightroom default setting includes some sharpening, these numbers are higher than the raw channel numbers that I have posted earlier. For now, you should compare these numbers only to the other ones in this test. 

The f/2.8 distribution is nice and tight. I haven’t tried yet, but I don’t think I could do that well with manual focus. However, the f/4 and f/5.6 distributions are by comparison pretty sloppy. Not only that, but the averages for those two stops are lower than the average for f/2.8. That’s not the way it should be. We have seen from the previous tests that the 63/2.8 lens is sharper at f/4 and f/5.6 than it is wide open.

I need to but this behavior into context. I’ve never tested an autofocus lens/camera system that is anywhere near as sharp as the GFX and the 63 mm lens. The sharper the imaging system, the more difficult is is to achieve a focus that gets the most out of it. SO I don’t know if there si a higher quality alternative available. This is making me want to test the X1D. 

At f/8, the spread is tightening up. That doesn’t necessarily mean that the AF is getting more accurate; it could be — and probably is — that he depth of filed is starting to make the foucs accuracy less important. That is proably also the explanation for the nice tight spreads at f/11 and f/16.

The average at f/4 and f/5.6 is not only lower than that at f/2.8, it is lower than that at f/8, where diffraction is holding down the on-axis sharpness. Thus, I think it’s safe to say that the lens is not achieving critical focus at f/4 and f/5.6. That makes me ask myself if it’s actually focusing at the taking aperture for those two stops. But the more important thing is that is does not seem to be constantly focusing in the same place.

Let me be clear that the average at any aperture never drops below 2000 cy/ph, and 2000 cy/ph is exceedingly sharp. I’m not saying that any of these sets of readings means that the camera’s AF system delivers fuzzy images: just not as sharp as they should be.

If we look just at one of the green raw channels we see roughly the same thing:

The numbers are slightly higher even though Lightroom sharpens during demosaicing.

Here’s the raw data for the Gr channel:

You’ll note that f/5.6 produces some sharp and some not-so-sharp results, depending on the luck of the draw.

Test conditions:

• ISO 100.
• 5000K illumination, bright enough so that the f/2.8 exposure was 1/320 second
• Target distance 3 meters
• AF focus/release/priority: focus
• Rapid AF: off
• Pre-AF: off
• AF-S
• 117 AF point mode, center point chosen, default size.
• EFCS on
• Pre-release half-shutter depression time: 750 milliseconds
• Shutter release full depression time: 250 milliseconds
• Camera moved 4 mm between exposures over a total range of 0 to 160 mm (40 moves, 41 exposures per stop)
• Camera motion applied by Cognisys focusing rail, with 1250 msec post-move settle time
• Shutter timing does by the same Cognisys controller
• RRS heavy-duty legs
• Arca-Swiss C1 head

The target was selected to be easy to focus on:

The above data is not the first series of test images that I made. Initially I made images with alternating 500-millisecond half presses and no half presses at all, with the camera set to release priority.

It’s clear from the above curve that 500 msec with release priority is not long enough ath f/8 and f/11. Interestingly enough, the results at f/4 and f/5.6 don’t seem to have been adversely affected.

With no half-press at all, just an (electronic) mashing of the shutter release:

That’s even worse.

I don’t normally show raw data, but the sequence of 500-msec half press exposures in instructive:

I’ve circled in red the points that I’d like to call to your attention. They are the f/4 and f/5.6 results that are high. It may be that the average is being dragged down not by a systematic focusing at the wrong f-stop, but by random focusing errors. 

This next graph is the raw data from the 750 msec, focus priorit series with which I led off:

It’s hard to say, but looking at this graph makes me think that there is some systemic source of error, since there are hardly any f/4 or f/5.6 poits as high as the bulk of the f/2.8 ones. You can also see that both f/4 and f/5.6 have some quite low values.