Panasonic Lumix GH5S sample gallery

But they got to that with last creation. I guess they figured that having camera supported by major SW from day one is better than rely on being "some day in future" supported.

Supports for camera RAWs should be done BEFORE its even released. So customers could go, buy, shoot in RAW and process as usually.

Sadly nothing works on new RW2. RT usually does, but not in this case..

LX100 is 3-4 years old, can not change lenses and instead features a fairly bright, big-ish zoom built-in. :)

In other words, DPR raves about low-light video with the GH5s, but I see the potential for stills too. If only that tech could trickle down to a less premium camera model.

What's your processing software? Kool-Aid room?

DPFan32, RX100's sensor suffers more.

liberty aint what you think: Makes sense...thanks.

I'll have to go back and try again on the A9. Whether photo or video, I haven't found a situation where I felt my A9 images were usable above 51,200.

For low light areas in museums or aquariums, I've actually had scattered A7s2 photos that I felt usable at 80,000 with plenty of video clips at ISO 102,400. Though I haven't personally replicated their same success, I've seen several decent looking A7s2 video clips up to ISO 204,800 and beyond...including Phillip Bloom's most recent comparison against the GH5s.

Karroly:

"Downsampling combines pixels, and combining pixels averages the noise."

Doing nothing to change the overall average noise of the image in the process. So, no, downsampling does NOT reduce noise as an overall percentage of the image.

"DPR has explained many times that the noise of an image, assuming the size of the "print" is the same, has virtually nothing to do with the pixel size but the sensor size. "

Print size is a different subject. Nothing to do with the point here. Now in the real world of ink to paper, one trick to reduce visible noise is to print many more dots per inch than say the normal 200-300.

"Just experiment and look."

Oh I have, I've tried to recreate the maths breaking rule, can't be done with basic Photoshop downsampling, no matter the pattern selected.

Don't make silly assumptions that I've not tried the trick. It doesn't work.

Karroly, continued:

Now, what may come to pass in the future, and already exists in DXO and PhotoNinja for raw files during extraction, is the capacity to compare adjacent pixels in tiff files. Do that say 4000 times per 100 pixels, then overlap those 100 pixels with the next 100, and the laws of averages will allow software to pick out what's a noisy pixel most of the time.

That process will take several minutes per tiff from a normal 24MP sensor. And it's not available as a "downsampling" option today in early 2018.

liberty aint what you think.
"So, no, downsampling does NOT reduce noise as an overall percentage of the image."
There is a difference between this theorical statement and what our eyes can see.
Let's say the noise can have a maximum value of +n and a minimum value of -n. The maximum noise difference between two adjacent pixels can be 2n. Let's assume I downsample a 24MP image to 6MP. A square of four pixels are combined into one pixel. If those four pixels have a noise value of +n, the resulting pixel will have a value of +n too. The same thing can happen with a group of four pixels with the -n value. So, for sure, on the downsampled image, it may happen that two adjacent pixels have a noise difference of 2n, too. BUT, the PROBABILTY of all the combined pixels having the same maximum, or minimum, noise value decreases when the number of combined pixel increases, because the noise is random, resulting in an image that looks less and less noisy...

liberty aint what you think, continued:
"Don't make silly assumptions that I've not tried the trick. It doesn't work."
Then I suggest you wipe off that dust on your computer screen... ;-)

Karroly:

"Then I suggest you wipe off that dust on your computer screen... ;-)"

I suggest you try the process, it doesn't work.

Also your math makes MY POINT for me, since this point of yours is absolutely untrue:

"noise value decreases when the number of combined pixel increases, because the noise is random,"

Random noise will still have the same average concentration in the image after downsampling; that's how the math, and downsampling, work.

Say you want to remove randomly distributed blue pixels at intensity X from a sample, you'll have to write code to find those pixels. A downsampling of the image won't do anything to exclude just the blue pixels of intensity X.

Or do noise thinking correctly: 500 pixels with 2 percent noise downsampled into 100 pixels will still have the same 2 percent noise--so that's 10 pixels for the bigger sample and 2 for the small smaller. There's no singling out of noise pixels in the downsampling, averaging, process you describe.

A old Canon 5D will make you happy than... Though no Video or other gimmicks

Take a pentax K1 and put it in apsc mode and you get big pixels.

liberty aint what you think,
"Say you want to remove randomly distributed blue pixels at intensity X from a sample, you'll have to write code to find those pixels. A downsampling of the image won't do anything to exclude just the blue pixels of intensity X."
When downsampling, that blue pixel will be merged with the surrounding pixels and the resulting pixel will be less blue...

liberty aint what you think, continued...
"Random noise will still have the same average concentration in the image after downsampling"
Wrong...
Merging changes the probability for one pixel to have a given noise level. Let's say the noise value can have 10 levels and is truly random. For each pixel, the probability of each noise level is 1 out of 10. It also means that, in a 1000-pixel picture, an average of 100 pixels (10%) will have the same noise value. Now, downsampling this picture by 4x will merge a square of 4 pixels into one. The total number of combinations is 10^4= 10000 (10 levels, 4 pixels combined). Among all theses combinations, several will produce the same noise level, but ONLY ONE produces the maximum noise level : all the four pixels must have the maximum noise value.
Thus, the probability of one pixel of the downsampled image to have the maximum noise level is now 1 out of 10000.... 1000 times lower than in the original image.

liberty aint what you think, continued...
Now, if you cannot see a significant noise reduction when downsampling a little bit (less or equal than 4x) an image coming for a Bayer sensor, especially in the blue and red channels, this is because of the Bayer filter structure. A 24MP sensor has 12MP green pixels, 6MP red and 6MP blue pixels. So, downsampling by 4x creates a new image of 6MP. It is easy to understand that in this case, the structure of the blue and red noise will not change. The noise specks will only be shrinked a little bit by the downsampling.

liberty aint what you think, continued...
Now, I made a little experiment for you. In my gallery, I have put three images. The first one is a 1000x1000-pixel, computer generated, black and white noisy image. The second one is a 4x downsampling of the first one, and the third is a 16x downsampling. There are here : https://www.dpreview.com/galleries/3570392623/albums/noise-tests
Please, view the pictures side-by-side at 100% magnification. If you cannot see that the downsampled images has less noise, then, for sure, you have got a problem...

Karroly:

"Merging changes the probability for one pixel to have a given noise level."

No, it doesn't, as an average of the total number of pixels. The downsampling does reduce the noise count, but also reduces the count of non-noise pixels.

And the non-reduction in overall noise is easy to see in your link.

In other words, the link shows no reduction in noise. Don't confuse NR with image size reduction.

Now, later today, I can try looking at the link on an only 8 bit monitor and then perhaps some of the colour noise, which is real easy to see on my monitor, will be greyed out more in the smaller images. Meaning what is real easy to see as colour noise will be hidden in grey on an 8 bit monitor.

As I suspected, your test shows exactly what I've said all along.

Because it's not like I've never tried the "downsampling" NR trick.

Karroly:

So I looked at your demo link on an iMac's 8 bit screen, and the noise is still the same average throughout the 3 samples. Now, it is less obvious, and looks more "grey", in all of the images, than on my 10 bit monitor, so that may be what's allowing you to think you're seeing an average reduction in noise with the downsampling.

The average stays the same.

Later, I can try an Macbook Air and an HP all-in-one with, both with lower quality 8 bit screens than the 8 bit iMac screen, and I imagine that I'll see the images as even more "greyed" out, so that will help to give the illusion of an average noise reduction.

Karroly:

Yep, tried a MacBook Air, and the images in your link look "greyer" than with the 8 bit iMac screen, but the average noise level remains the same in all 3.

liberty aint what you think,
"No, it doesn't, as an average of the total number of pixels. The downsampling does reduce the noise count, but also reduces the count of non-noise pixels."
Obvioulsy, you did no understand my demonstration, so I will make it more simple.
Let's say we have five noise levels (-2, -1, 0 , +1, +2). Probability of each level 1/5 for a truly random noise. Downsampling of 2x to merge 2 pixels into one.
This make 5^2 = 25 possible combinations for the noise level, detailed as this :
+2 = (+2,+2 ) = 1
+1.5 = (+2,+1) (+1,+2) = 2
+1= (+1,+1) (+2, 0) (0,+2) =3
+0.5 = (+1, 0) (0,+1) (+2,-1) (-1,+2) = 4
0 = (0, 0) (+1,-1) (-1,+1) (+2,-2) (-2,+2) = 5
As this is symetrical :
-0.5 = 4
-1 = 3
-1.5 = 2
-2 = 1
1+2+3+4+5+4+3+2+1 = 25
As noise values are binary integers, after rounding :
+2 = 1
+1 = 5
0 = 13
-1 = 5
-2 = 1
With 1+5+13+5+1=25
New probabilities of noise levels :
+2 or -2 = 1/25
1 or -1 = 5/25 = 1/5
0 = 13/25
Conclusion : picture less noisy.

Even if I do not round the new half values, here is the comparison between the noise distribution of the original picture :

1/5 1/5 1/5 1/5 1/5
-2 ...-1...0... +1...+2
(This is an horizontal flat curve.)
and the downsampled picture, no rounding :

1/25 2/25 3/25 4/25 5/25 4/25 3/25 2/25 1/25
-2.....-1.5...-1....-0.5....0...+0.5...+1...+1.5...+2
This is a triangular shape. With a higher percentage of pixels having a low noise level.

J A C S,
There is no free lunch ;-)

@JACS - Well it's both, isn't it?

Comparing noise on images with a different pixel count is not such a great idea in the first place. You may want to compare them at the same size. If that size is smaller than the smaller image, the noise will be the same. If it is larger, then the downsized image will have less noise because the high frequency component will be lost.

J A C S,
For sure. But the discussion here was about the effect - or lack of effect - of downsampling on noise, without considering the "collateral damages"

It does not make much sense to compare noise on two images of different pixel count, that is my point. Noise is not just one number - it has a spectrum, as well, for example. If you change the pixel count, you get a different animal.

Karroly,

Just from your 3 image sample there is no average reduction in noise as you downsampled. None, and it's easy to see this non-change.

"'No, it doesn't, as an average of the total number of pixels. The downsampling does reduce the noise count, but also reduces the count of non-noise pixels.'"

I suggest you look at your own examples to verify my point.

"Let's say we have five noise levels (-2, -1, 0 , +1, +2). Probability of each level 1/5 for a truly random noise. Downsampling of 2x to merge 2 pixels into one.
This make 5^2 = 25 possible combinations for the noise level, detailed as this :"

You're not merging levels, that takes processing, and time, a good deal of it. You're averaging the whole image, that's why the process is basically instant.

The appalling thing is your refusal to look at your own examples.

"Noise is not just one number - it has a spectrum, as well, for example."

Nope noise is just a number, or average, and downsampling doesn't change that number.

So if downsampling doesn't work, why does the 42MP A7r II look as good - if not better - at high ISOs than the 12MP A7s II when viewed at the same output size?

Is it pure magic that Sony has managed to produce pixels 3.5x smaller with zero degradation in performance whatsoever?

Or is it instead that pixel count is basically irrelevant to noise performance, and sensor size and output size are the factors that actually matter?

Androole:

The A7RIII is a few generations of CMSO technology in advance of the A7SII, which is not an especially strong higher ISO body in 2017 terms.

The A7RIII is also BSI.

Both the Nikon D850 and Sony A9 are better higher ISO bodies than the A7SII.

"Or is it instead that pixel count is basically irrelevant to noise performance,"

Capacity to disperse heat is the most important thing.

Jeez, you're trying to be persnickety in order to evade the fundamental point.

Let's get back to basics, then.

https://www.dpreview.com/reviews/image-comparison?attr18=lowlight&attr13_0=sony_a7r&attr13_1=sony_a7rii&attr13_2=sony_a7s&attr13_3=sony_a7sii&attr15_0=raw&attr15_1=raw&attr15_2=raw&attr15_3=raw&attr16_0=6400&attr16_1=6400&attr16_2=6400&attr16_3=6400&attr171_1=1&normalization=compare&widget=1&x=-0.880742502681778&y=-0.18644523925272583

That's the A7r, A7r II, A7s, and A7s II. The A7r is the oldest camera for a good margin, is not BSI, has 3x the MP count of the A7s, and still looks very competitive at high ISO...but only when downsampled to the same resolution.

Time travelling CMOS magic?

And if you're talking about heat-dispersal, the solid-mounted sensor of the A7s and A7r are both in a better position to achieve that than the IBIS systems on the A7r II and A7s II. And yet the A7r II looks the best of all of them, and the two A7s cameras are functionally identical...

i am happy with 2mpx dash cam better low noise video :D

The design of a GH5S is designed like every other DSLR-like stills camera.
The GH5S is not MENT to be a stills camera, you could say.
But it's designed that way like a stills camera :) Let's be correct :)