I have previously reported that the Sony a7RII performs low-pass spatial filtering on raw files under some conditions: bulb and high-ISO exposures. This behavior is now often called the “star-eater” algorithm, a reference to its degradation of some astro images. Recently, an astrophotographer posted some images here, and claimed that the star-eating behavior affected timed exposures of 4 seconds and above.
This report was greeted with widespread skepticism, including some from me. However, the photographer kept producing images that seemed to hint that the effect was real. I had done my original testing shortly after the a7RII was first shipped using firmware version 1.0. The current version is 3.30. Could it be that the camera’s behavior has changed?
I retested the camera with FW 3.3, and found that out. I will now present my findings. If you are a skeptic — as I was — I hope this convinces you.
Here are spectra from raw red channel dark-field exposures at 3.2 and 4 seconds. Take a look, then read on for my explanation:
First, the conditions under which I made the exposures:
- No lens on camera
- Uncompressed raw (compressed gives similar results)
- EFCS on (not that it matters)
- Mechanical second curtain
- Single shot shutter mode (this does matter).
fs is the sampling frequency of the sensor. f/fs = 0.5 is thus the Nyquist limit.
Dark field exposures are images of the read noise of the camera. Read noise is divided into two components: that part that is correlated from frame to frame, and that part that is uncorrelated. The portion that is correlated tends to have higher energy at low frequencies. The part that is uncorrelated should be what engineers call white, that is, have equal energy at all frequencies.
If the read noise were truly white, the charts of the read noise would be flat from the left side of the above graphs to the right. If you look at the top graph, that is the case for the vertical frequency components. Such flatness indicates no spatial filtering on the part of the camera firmware. The horizontal component shows a roller-coaster periodic behavior. I am not entirely sure as to its origin, but I saw similar behavior at some ISO settings with FW 1.0, I believe it is related to differences in the column ADCs and column amplifiers on the a7RII sensor.
In the 4-second image, the vertical curve is no longer flat, but hows a 1 db droop from low frequencies to the Nyquist frequency. That is the result of the camera’s firmware. It didn’t happen with FW 1.0.
The spectra that happen at 4 seconds continue all the way up to a 30 second timed exposure:
However, with a bulb exposure, things are different:
When bulb is used, there is less spatial filtering of the vertical components, and, once you mentally subtract off the peaks of the horizontal component and look only at the troughs, about the same amount of filtering of the horizontal component.
While I was conducting the testing that lead to the above graphs, I noticed some anomalies in the histograms of the FW 3.3 camera. I’ll be tracking those down and reporting on them soon. Also, I’ll be exploring other ISOs than the base ISO. Unfortunately, I have upgraded (although if I were an astrophotographer, I’d argue that it wasn’t an upgrade) all my a7RII’s to FW 3.3, so I won’t be able to do side by side comparisons.
Fortunately, I did fairly extensive spectral testing of the original firmware, and I’ll give pointers to what I posted about that. If you want to see how the old firmware performed in tests similar to those posted today, take a look here and here.