Fuji GFX50R and star eater

Started Jan 27, 2020 | Discussions thread
Flat view
weihaowang Forum Member • Posts: 78
Fuji GFX50R and star eater

I have been a DPR member for many years, but I seldom posted on the forums here. However, I realize that DRP is where most of the "star eater" discussion occur, I feel it's appropriate to share my findings here and ask for feedbacks.

The end of last year, Santa brought me a Fuji GFX50R and a couple of lenses, and that's when the headaches started. Mark helped me to find that the dark files of the 50R seem to have some filtering. So I spent some time under the real sky to figure out what's going on.

I will present all the test results, but here let me just say that Fuji seems to do hot pixel suppression (plus perhaps smoothing of image at the same time) on raw files from the 50R, at least on long exposure images.  So be careful if you are picky and are thinking about getting a 50R (perhaps also the 50S and the 100) for astrophotography.

(Let me also say that Long Exposure Noise Reduction doesn't seem to turn these off, so we can't simply get real raw images with LENR-on even if we are willing to sacrifice 1/2 of our imaging time. I also tried the mode-3 method, which was used to fight against Nikon's star eater in the old days, and the results are not conclusive. I will try this again.)

Now the long story starts. You can just read the previous paragraph and stop there if you are not interested in the details.

As soon as I got my first deep stacking of 50R images, I noticed that the stars don't look right. The faint stars look very colorless (comparing to brighter stars) and soft, although all other evidence shows that the lens (GF110/F2) is extraordinarily sharp. Not only the softness, they just don't look right.

In the image below, you can compare the two in the right (the two stacks of 20 and 22 50R images) against the one in the left (modified D800 with firmware hack, so no raw cooking at all). I believe you can see what I meant by faint stars not looking right. To be more specific, in the reference images taken by D800, you can see stars of full ranges of brightness, from very bright to very faint and everywhere between these two extremes. On the other hand, in the 50R images, you can only find bright stars and very faint stars, the stars of intermediate brightness are gone. Well, they are not gone, they just become fainter. All the faint stars in a deep 50R image are about equally faint.

GFX50R vs hacked D800

(Another thing that worries me is that although the 4min*22 50R image has longer integration time than the D800 image, it's much noisier. However, I can't say for sure whether this means some problem in the camera performance. This can simply be because the 50R is not modified while the D800 is. This can also be because the Sigma lens used on D800 has larger aperture diameter than the 110mm lens at F2.8 (but let's not fall into the fight of F-number vs aperture here). There are other images I took that also suggest 50R is a noisier camera, but I don't think it's conclusive enough yet.)

Back to the faint stars. The next two images shows different stacks of 50R images of 1-minute subs (ISO 3200) and 4-minute subs (ISO800). The images are naturally dithered (by the mount's tracking error, and therefore the 4-min ones have slightly elongated stars). They are stacked in PixIinsight with outlier rejection, so hot pixels (if any) are gone.

Various stacks of 1-minute subs.

Various stacks of 4-minute subs.

In both of them, if you look at the stacks of one or four subs, you can find many colorful stars, kind of too colorful. You can also find some stars with "holes" in the middle. I believe these are strong indications of hot pixel suppression, like those in Sony and some Nikon models. These colorful stars turn to have more natural colors once the numbers of stacked images increase to 8 and higher, as the dithering helps to average out the effect of losing one or two pixels in a star.

Then let's turn our attention to the faint stars again. In the above two comparisons, if we look at the stacks of 20 or 24 subs, we can clearly see the many soft and colorless faint stars I mentioned earlier. These faint stars are fainter than the noise level in a single sub, so you can't see them in the stacks of one or four subs. They only start to show up when the number of stacked subs is large enough, so the overall noise level becomes lower than the faint stars. Roughly speaking, you can start to see them in the stacks of 12 subs. However, here they are still not too much brighter than the noise level in the stacked image, so you won't notice their "strangeness." Only after stacking more than 16 subs, their S/N becomes sufficiently good and their strange appearance becomes noticeable. Note that the threshold when strange-looking faint stars start to appear is the number of stacked subs, not the total integration time, as clearly shown by the above two examples.

I am not sure whether it is the same mechanism that produces the color and holes of bright stars also makes faint stars soft and colorless. It may be the same hot pixel suppression algorithm. Or, perhaps the faint stars are suppressed by another mechanism, since they are indistinguishable from noise in single subs. Nevertheless, by looking at the stars produced by the 50R, especially by comparing them with D800, I am very certain that 50R's raw files are cooked badly. But this is not good enough. If I am to write a letter to Fuji and ask them to correct this, I need stronger evidence. I need something that's so obvious that even a untrained eye can easily see the difference. The D800-50R comparison is close, but not there yet, since the sensors have different pixel sizes and the lenses are also different.

Unfortunately, I haven't obtain good comparison images taken with the same optics and the same sensor. I tried to adapt a Pentax 645 55mm lens to 50R and also put it on Pentax 645z. Unfortunately the lens adapter keeps the lens at its smallest aperture (F22) on the 50R. This puts the image in the dark+read-noise-dominated regime, and any small difference between 645z's and 50R's dark and read noise levels can ruin the comparison. I also tried both cameras on TAK TOA-150 and image the cluster M44. It's an F7 scope, not as bad as F22. However, under its 1100mm focal length, seeing dominates the star size so stars are not as sharp as they are under short camera lenses. We know that if the stars are not sharp, then the filtering effect will not be very apparent. So the comparison images I took with the TOA cannot be used in my letter to Fuji.  But anyway, we can still learn something from such images.

645z and 50R under TAK TOA-150, unstretched images

645z and 50R under TAK TOA-150, stretched images

You can first look at the unstretch images (the top one). You will see that the intermediate-brightness and faint stars in the 50R images look much greener than in the 645z images. I cannot make this green tint go away by simply changing the color balance, and I will come back to this point later. Anyway, the green tint is a somewhat subtle indication that 50R did something to the stars.

If we look at the stretched images very very carefully, we see that some faintest stars/galaxies in the 645z images are gone or become fainter in the 50R images. There are counter examples, but I think overall, 50R loses more stars than 645z does. This effect is again quite subtle. However, unlike the comparison made on the unstretched images, where the stars are all well above the noise level, here we are looking at stars quite close to the noise level and therefore the detection limit of the images. So the differences we pick up here could be cause by some raw cooking made by 50R, or simply by the difference in sensitivities (for example, more dark noise in 50R, or less light throughput).

I think what we can conclude here is that once used on telescopes of long focal lengths, the behavior of 50R is much more acceptable. We can still find something not quite right in a 50R image, but that's only when we have a golden standard to compare with. Without such a comparison, 50R images can look just completely normal to most people.

Now back to the green tint in the stars in the unstretched 50R images. I want to make sure the green tint is not caused by difference in color balance. What I did is to divide the 14-stacked 50R image by the 15-stacked 645z image. If the green tint is just a color balance issue, then all stars should have the same ratio in the divided "ratio image" in each channel.  But it turns out this is not the case.  Stars of different brightness have different ratios. This is shown in the diagrams below.

50R/645z brightness ratio on stars in M44

The three diagrams show the 50R/645z pixel brightness ratios as functions of pixel brightness, for the R, G, and B channels. You can see that in all three plots, there are two horizontal (or slightly tilted) sequences. Most of the stars belong to the lower sequence in G and B, while most of the stars belong to the upper sequence in R.  A simple way to look at the existence of the lower sequence is that stars on this sequence have suppressed brightness in the 50R image, comparing to stars on the upper sequence. The real question is, what are those stars on the upper sequence and lower sequence?

I further tried to identify pixels that belong to the upper and lower sequences, and check where they are in the image. In all the R, G, and B cases, the lower sequences correspond to intermediate-brightness stars, and the upper sequences correspond to the brighter and faint stars (quite interesting, right?).  Note that here we are looking at images stacked from 14 and 15 subs. So the faint stars in these stacked images should be embedded in noise in the single subs.  The existence of two sequences in the diagrams implies that 50R's raw cooking algorithm behaves differently on stars of different brightness.

Here, one might ask how can I be certain that it's 50R who is cooking the raw files, not 645z?  I can't be 100% certain.  However, I have used 645z for almost five years and I had made many extremely deep stacks.  I have never noticed anything strange on my 645z images, and all my 645z images are qualitatively similar to images produced by my hacked D800. So I have high confidence on the "raw-ness" of 645z files, although theoretically, I cannot completely rule out the possibility that 645z also does something to the raw files.

OK, the story stops here. Forecast says that the sky will become clear again in three or four days. I will find a lens with an aperture ring that can be used on both 645z and 50R. Hopefully I will get super convincing images to tell Fuji that their 50R is very naughty on astronomical images.  Now I will listen to suggestions and comments from you guys.


Nikon D800 Pentax 645Z
If you believe there are incorrect tags, please send us this post using our feedback form.
Flat view
Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow