This page takes a detailed look at frame averaging and how it can improve pixel quality with the Sony A7R IV at ISO 100. Conclusions and tips are included for those looking to raise their game on the image quality front.
This meticulous study addresses both pixel shift and single-shot multiple exposures. The findings here apply in general to any camera, excepting of course Sony specific behavior, which is noted.
Includes detailed analysis and crops.
While noise with the Sony A7R IV is well controlled, there are very large benefits to be gained with some images, particularly those with a wide dynamic range, and with minimal effort. For those looking to make very large prints, the extra effort (not much) can mean being able to sharpen more and scale up more with minimal artifacts becoming noticeable nearly as soon.
Camera vendors lack creativity and imagination!
Apple is 'killing it'. Camera vendors are not. They deserve to go out of business soon for their total failure of imagination and creativity in making cameras automate stupidly easy stuff for me and for you that could greatly improve image quality. Exception: Panasonic S1R with its Multi-Shot High-Res mode, and PhaseOne, with frame averaging built in.
Consider the billions of dollars spent to get 1/3 or 1/2 stop more dynamic range and a little more resolution when software can yield 2/3/4 stops more, by a firwmare update. It’s probably 100X simpler than doing autofocus well—shoot and average frames, and output a single 16-bit raw file. See comments further below .
Vincent V writes:
I appreciate you image averagin techniques, but there may be a better way to get noiseless photography, using only 4 frames http://www.guillermoluijk.com/article/nonoise/index_en.htm
DIGLLOYD: there are several undeniably better ways to do frame averaging, but for me, the above link is useless : I use a Mac, so Linux and Windows software is a non-starter. Moreover, bracketing has its issues with real lenses in the field (flare and lens dynamic range for starters).
Indeed, the first thing I tried before frame averaging was exposure bracketing followed by Merge to HDR in Photoshop. I was not successful in getting a satisfactory result that looked right to me.
Camera vendors lack creativity and imagination!
Apple is 'killing it'. Camera vendors are not. Consider the billions of dollars spent to get 1/3 or 1/2 stop more dynamic range when software can yield 2/3/4 stops more, trivially!
It all ought to be a one-button press automatic process: the camera would shoot N frames automatically, averaging them all into a single 16-bit raw files, bracketing slightly if desired.
The only camera to do this well is the Panasonic S1R with its Multi-Shot High-Res mode, which has the awsesome bonus of oversampling for much higher resolution on top of the 2.8X reduction in noise. The PhaseOne IQ150 explicitly has frame averaging, but it’s way above my pay grade and relatiely slow compared to what a mirrorless camera could.
Thing is, the Sony A7R IV should be able to bang out 10 frames and merge them into a single raw file in just one second. Why is that feature missing? And why is it missing for Sony, Canon, Fujifilm? Nikon has it, but it is done poorly on the Nikon D850 at least (shutter bang for starters).
Nick M writes:
I have been enjoying your recent explorations of pixel shift and frame averaging. It prompted me to do a bit of experimentation myself with my A7RIII. I took two pixel shift sequences of the same scene and produced 4 outputs:
- Single capture image with sharpening
- Frame average of 4 single captures with sharpening after averaging
- Pixel shift image from single 4 image sequence
- Frame average of 4 pixel shift images each from 4 image sequence The scene had plenty of moving grass and lighting changes. All raw captures were given some quite aggressive contrast control before any further processing (shadows +100, highlights -50, contrast +25)
My key takeaways.
- Best image overall (to my eye) was the frame average of 4 individual captures. At 100 ISO I see little/no noise benefit versus a single capture, but this image was able to take substantially more sharpening and hence the finished image appears materially sharper than an individual capture. I did bring in a single capture layer to paint out some areas of ugly detail from the averaging process where there was movement.
- Single pixel shift image unusable due to checkeboarding from movement and lighting changes.
- Averaged pixel shift image much more usable in terms of checkerboarding, although still visible in places. Detail strong, but not better than 4 image average. Perhaps this conclusion would differ with better technique and/or certain subjects
- Single capture strong, with minimal noise despite the aggressive contrast control. For most purposes this would be a good output with which I would be happy. But for a large print the 4 frame average clearly shows more detail. Putting together the 4 frame average is reasonably time efficient (even with my slow machine), but painting in detail from a single capture where there is movement is time consuming.
A 4 frame average was put together from the single captures for one of the pixel shift sequences (after aligning them of course). This makes me think that a potential field workflow is to make a pixel shift capture (or multiple such captures) since this allows for multiple post-processing solutions if desired with a frame average being the likely preferred output in most landscape scenarios.
I also took a quick look at some high ISO files and my initial impression is that the noise reduction benefits from averaging can be compelling. More testing certainly needed.
I would be interested to know where in the process you do sharpening. I decided to do contrast control, but no sharpening to the raw files before averaging. I certainly haven’t done exhaustive testing, but common sense seemed to suggest this as the best workflow and I did find the averaged file to be far more tolerant of sharpening than individual captures.
DIGLLOYD: sounds about right. I have been sharpening before averaging. I don’t think no sharpening is a good idea, but less sharpening during raw conversion followed by aggressive sharpening after averaging possibly would produce better results. Partly it depends on the tools before/after and how the sharpening works.
Glenn K writes:
The laws of S/N ratio apply just like in seismology. S/N ratio improves by the square-root of the number of shots. So two shots gives you 1.4 better S/N, 4 shots doubles S/N, but then you need to go to 16 shots to double again... so, as you show, 2-4 shots is the sweet spot.
DIGLLOYD: absolutely... but the series I show looks better than 1.4X by going to 2 frame average... a factor of 1.4X seems much less than it appears to my eyes (2 frames). I am guessing there is more to it, e.g., some averaging out of the randomness in a beneficial way, not just S/N ratio. Or, perhaps it is just a perceptual trick, but that after all is what counts, not the math, since the brain has all sorts of weird tricks in doing visual processing.
Ken L writes:
Thanks for all the helpful content you have been sharing lately: outstanding ! Perhaps I have overlooked it, but have you given any recommended method of frame averaging ? As a Photoshop user I found this illustration on https://www.cambridgeincolour.com/tutorials/image-averaging-noise.htm but it may be outdated.
DIGLOYD: I do have a recommended method, and I will probably offer a script to make it super easy from a set of TIF. But I'm heading out on a trip, and that will have to wait. But in essence, it would automate what that link above does, saving time and risk of error.