k1 pixel shift - can you see the different?

Pixel shift of course may have an effect on details, but that depends a lot on overall sharpness of the scene (lens vs sensor etc).

But pixel shift has a definitive effect on reducing noise, as you take the mean of 4 captures instead of just 1.

In an evenly lit scene like this it's not really perceivable, but even at ISO 100 in a scene with dark areas it can have a very noticeable impact. For example I used it when shooting boats at night in the harbor, strong contrast between lights and very dark areas, and pixel shift greatly improved the IQ of my images !

JeremieB wrote:

Pixel shift of course may have an effect on details, but that depends a lot on overall sharpness of the scene (lens vs sensor etc).

But pixel shift has a definitive effect on reducing noise, as you take the mean of 4 captures instead of just 1.

I have never understood HOW this works. I've done tests to show that it DOES work.

If you have PS off, a pixel is might be Red, Green or Blue.  The other colours are made up the average of the neighbours with that colour.

But you have PS on there is a single reading for Red and Blue and two for green each site.  There should be less averaging with PS.  (Unless each image is demozaiced and the 4 are averaged, but that would reduce resolution)

There's more light captured, so there should be less noise, but I can't get my head round the mechanism, unless it the extra data playing a role in NR.

I didn’t read all the responses, so someone else may have already mentioned this. The images say they were 2.5 second exposures. You’re kidding yourself. If you think a plant will be totally motionless for 2.5 seconds. Use something printed on paper for a test object. I often use the outside of a shipping container that has a barcode and other minute details. And give it more light.

James O'Neill wrote:

JeremieB wrote:

Pixel shift of course may have an effect on details, but that depends a lot on overall sharpness of the scene (lens vs sensor etc).

But pixel shift has a definitive effect on reducing noise, as you take the mean of 4 captures instead of just 1.

I have never understood HOW this works. I've done tests to show that it DOES work.

If you have PS off, a pixel is might be Red, Green or Blue. The other colours are made up the average of the neighbours with that colour.

But you have PS on there is a single reading for Red and Blue and two for green each site. There should be less averaging with PS. (Unless each image is demozaiced and the 4 are averaged, but that would reduce resolution)

There's more light captured, so there should be less noise, but I can't get my head round the mechanism, unless it the extra data playing a role in NR.

Take 1 picture of size m x n, resize in half to get a picture of size m/2 x n/2 (remove high frequencies before resizing of course to avoid aliasing).

New picture will have better S/N ratio, it's sort of pixel binning in fact (half pixels with 4 times more "light" per pixel). That's the averaging I talk about. Averaging being a blur - there's less visible noise and better S/N. That's also exactly the same effect as when we capture a huge number of photos in astro, then stack them to reduce noise. Averaged same signal remains same signal, only the noise gets step by step lower / closer to the "true" signal (well, shot noise, noise that is not random is not concerned).

PS works the same except it starts from 4 images (that can be considered the same as 1 image of twice the size in each dimension, That's what others do with super resolution). Pentax PS prefers binning to produce an image with less noise, and with added value of cancelling the Bayer filter by moving the sensor appropriately. Of course you don't need PS just to lower noise with this technique, as I noticed about astro for example. But that + Bayer cancelling makes it a neat trick.

Of course PS effect might seem subtle or even useless compared to state of the art denoising and sharpening programs, but it's another approach / philosophy that is interesting too, and if you absolutely don't want false details AI software is not the best idea

JeremieB wrote:

James O'Neill wrote:

JeremieB wrote:

Pixel shift of course may have an effect on details, but that depends a lot on overall sharpness of the scene (lens vs sensor etc).

But pixel shift has a definitive effect on reducing noise, as you take the mean of 4 captures instead of just 1.

I have never understood HOW this works. I've done tests to show that it DOES work.

If you have PS off, a pixel is might be Red, Green or Blue. The other colours are made up the average of the neighbours with that colour.

But you have PS on there is a single reading for Red and Blue and two for green each site. There should be less averaging with PS. (Unless each image is demozaiced and the 4 are averaged, but that would reduce resolution)

There's more light captured, so there should be less noise, but I can't get my head round the mechanism, unless it the extra data playing a role in NR.

Take 1 picture of size m x n, resize in half to get a picture of size m/2 x n/2 (remove high frequencies before resizing of course to avoid aliasing).

New picture will have better S/N ratio, it's sort of pixel binning in fact (half pixels with 4 times more "light" per pixel). That's the averaging I talk about. Averaging being a blur - there's less visible noise and better S/N. That's also exactly the same effect as when we capture a huge number of photos in astro, then stack them to reduce noise. Averaged same signal remains same signal, only the noise gets step by step lower / closer to the "true" signal (well, shot noise, noise that is not random is not concerned).

That's all good.

PS works the same except it starts from 4 images (that can be considered the same as 1 image of twice the size in each dimension, That's what others do with super resolution).

Argh. Typing up why that's not right , I might have understood why it is. I need to go away and do the sums on what happens if you project a squares 1,2, 3 & 4 pixel widths of white or (e.g.) red light offset left/right by 1/4 pixel and up/down but 1/4 pixel.

Possibly my head will explode before I get to the answer.

James O'Neill wrote:

JeremieB wrote:

James O'Neill wrote:

JeremieB wrote:

Pixel shift of course may have an effect on details, but that depends a lot on overall sharpness of the scene (lens vs sensor etc).

But pixel shift has a definitive effect on reducing noise, as you take the mean of 4 captures instead of just 1.

I have never understood HOW this works. I've done tests to show that it DOES work.

If you have PS off, a pixel is might be Red, Green or Blue. The other colours are made up the average of the neighbours with that colour.

But you have PS on there is a single reading for Red and Blue and two for green each site. There should be less averaging with PS. (Unless each image is demozaiced and the 4 are averaged, but that would reduce resolution)

There's more light captured, so there should be less noise, but I can't get my head round the mechanism, unless it the extra data playing a role in NR.

Take 1 picture of size m x n, resize in half to get a picture of size m/2 x n/2 (remove high frequencies before resizing of course to avoid aliasing).

New picture will have better S/N ratio, it's sort of pixel binning in fact (half pixels with 4 times more "light" per pixel). That's the averaging I talk about. Averaging being a blur - there's less visible noise and better S/N. That's also exactly the same effect as when we capture a huge number of photos in astro, then stack them to reduce noise. Averaged same signal remains same signal, only the noise gets step by step lower / closer to the "true" signal (well, shot noise, noise that is not random is not concerned).

That's all good.

PS works the same except it starts from 4 images (that can be considered the same as 1 image of twice the size in each dimension, That's what others do with super resolution).

Argh. Typing up why that's not right , I might have understood why it is. I need to go away and do the sums on what happens if you project a squares 1,2, 3 & 4 pixel widths of white or (e.g.) red light offset left/right by 1/4 pixel and up/down but 1/4 pixel.

Possibly my head will explode before I get to the answer.

ok you're right it's not exactly the same as same image taken by a sensor with 4 times more pixels, or let's say it depends how much the smaller sensor moves between each shot.

BUT what matters for the effect on noise is that each square group of 4 pixels (if I take same colored pixel from each shot and put them side by side), once averaged together, represent the average of the signal under it - so, less noise

Of course if I move the sensor more than 1 pixel between shots then the 4 resulting pixels (4 R, 4 G x 2, 4 B), will not be the average of the underlying signal anymore.