Reading mode: Light Dark

Noise... camera ISO sweet spots?

  • Thread starter Thread starter Ubilam
  • Start date Start date
Go to last post
U

Ubilam

Guest
I see something that seems to be ignored here. The same camera can have more noise just a step or two up and then its clean again...ie: iso 100 can be as clean 320. 125 as bad as 400. Some high ISOs near to each other vary too more or less. Whats up with that?

Ubi.
 
It sounds like you're seeing the effects of noise reduction being applied at different levels at different ISO settings. You would only see that in JPEG. Most RAW files have no noise reduction applied ( although Sony, in particular, are known to process some RAW files and teh Nikon 1-series seem to have processing applied to RAW ).
 
On my Pentax K-30 you can adjust the JPG noise reduction to any of 4 levels (one being OFF) in 9 different ISO ranges; so you know what you're getting.
 
Member said:
Ubilam wrote:

I see something that seems to be ignored here. The same camera can have more noise just a step or two up and then its clean again...ie: iso 100 can be as clean 320. 125 as bad as 400. Some high ISOs near to each other vary too more or less. Whats up with that?
Click to expand...
...is the ISO setting that gives the desired image brightness for the desired f-ratio and shutter speed. There are two primary sources of noise in a photo:
  1. Noise from the light itself (photon noise)
  2. Additional noise added by the sensor and supporting hardware (read noise)
The more light that falls on the sensor during the exposure, the lower the photon noise. For cameras with ISOless sensors, the read noise is constant throughout the ISO range. For cameras with non-ISOless sensors, the read noise lessens with higher ISO settings.

For example, the Nikon D800 has an ISOless sensor. A pic of a scene at f/2.8 1/100 ISO 1600 will have the same noise as a pic of the same scene at f/2.8 1/100 ISO 100 pushed 4 stops in the RAW conversion, because the read noise is the same regardless of the ISO setting.

On the other hand, the Canon 5D3 does not have an ISOless sensor. A pic of a scene at f/2.8 1/100 ISO 1600 will have less noise than a pic of the same scene at f/2.8 1/100 ISO 100 pushed 4 stops in the RAW conversion, because the read noise is less at ISO 1600 than at ISO 100.

However, in both cases, a pic at f/2.8 1/100 ISO 1600 will be less noisy than a pic at f/2.8 1/200 ISO 3200, because twice as much light falls on the sensor for the ISO 1600 pic, and the photon noise matters more than the read noise.

So, no, ISO 320 will not be as clean as ISO 100, and ISO 125 will not be as bad as ISO 400, unless unusual processing is being applied to the photo, such as different amounts of NR (noise reduction) for OOC jpgs.

The main thing, then, is to get as much light on the sensor as you can within your DOF (aperture) and motion blur (shutter speed) constraints. This will result in the least amount of noise in the captured photo. Next, use the highest ISO you can within the constraint of blown highlights to reduce the amount of noise for a camera using a non-ISOless sensor.

And that's basically how it goes.
 
17e842a696954dfba1cb514a705eec54.jpg


this line might be a sweet point. it's *** dB/LM?
 
The D800(E) sensor is not ISOless. If it was, then we would see the DXOmark dynamic range plot drop by 1 Ev for each stop increase in ISO. In fact it drops only about 1/2 Ev from 100 to 200 ISO and drops 2.7 Ev instead of 3 Ev from 200 to 1600 ISO. Above 1600 ISO the D800(E) is ISOless. Another way to check for an ISOless sensor is to check the read noise in e- derived from the full SNR data. For the D800E that is the noise n in the following table (m is full well count and g% is PRNU):

NikonD800E

ISO: 100 200 400 800 1600 3200 6400 12800 25600

m = 54924 26293 13414 6722 3483 1776 833 438 237

n = 5.16 3.25 3.27 2.95 2.72 2.78 2.41 2.81 2.83

g% = 0.30 0.37 0.42 0.50 0.51 0.74 0.00 0.49 0.00

Note: Data at 6400 and 25600 ISO is unreliable because of implausible g% value.

We see the read noise starts at 5.16 e- then drops by about 1.9 e- at 200 ISO. Then it drops about another 0.5 e- to about 2.75 e- at 1600 ISO and above.

Compare this to the Canon 1DX table below:

Canon1DX

ISO: 100 200 400 800 1600 3200 6400 12800 25600 51200 102400 204800

m = 90101 46728 23720 11969 6296 3072 1524 738 332 163 104 52

n = 38.51 19.23 9.98 5.28 3.58 2.19 1.89 1.62 1.27 1.70 1.65 1.17

g% = 0.44 0.43 0.52 0.59 0.71 0.83 0.84 0.76 -0.00 -0.00 5.92 8.38

Note: Data above 12800 ISO is unreliable because of implausible g% value.

Here the read noise starts much higher at 38.5 e- at base ISO, then steadily drops to about 1.6 e- at 12800 ISO. So ISO gain does have a much greater effect on read noise for this Canon camera than for the Nikon D800(E).
 
Last edited:
Don't have a clue about 'sweet spots' in different iso settings and not much on the technical aspects or charts but have found this from shooting a lot of low light images.A properly focused and exposed image from the same camera can be acceptable even at iso3200 but show noise at iso 400 that is underexposed.It is hard to keep noise out of the shadows even in a properly exposed pictures.I have took nice photos of bees and such at iso 3200 that showed no noise in the bee but the out of focus background had plenty.If they are 'sweet spots' it would be helpful to find them.
 
It certainly is ...that there are different amts of noise shown at various isos, and it is not always a linear chart/increase/decrease. Just read published reports in magazines ...various cams, various mags. Why ? donno. But one would think it has to do with how the cam processes the light coming into it. Not linear.
 
Brad99 said:
Brad99 wrote:

On my Pentax K-30 you can adjust the JPG noise reduction to any of 4 levels (one being OFF) in 9 different ISO ranges; so you know what you're getting.
Click to expand...
I turn off all noise reduction in my 7D because I want to decide what is noise (and details) remain later in post editing. Letting the camera decide what is noise is not always a good idea.


Ubilam.
 
DSPographer said:
DSPographer wrote:

The D800(E) sensor is not ISOless. If it was, then we would see the DXOmark dynamic range plot drop by 1 Ev for each stop increase in ISO. In fact it drops only about 1/2 Ev from 100 to 200 ISO and drops 2.7 Ev instead of 3 Ev from 200 to 1600 ISO. Above 1600 ISO the D800(E) is ISOless. Another way to check for an ISOless sensor is to check the read noise in e- derived from the full SNR data. For the D800E that is the noise n in the following table (m is full well count and g% is PRNU):

NikonD800E

ISO: 100 200 400 800 1600 3200 6400 12800 25600

m = 54924 26293 13414 6722 3483 1776 833 438 237

n = 5.16 3.25 3.27 2.95 2.72 2.78 2.41 2.81 2.83

g% = 0.30 0.37 0.42 0.50 0.51 0.74 0.00 0.49 0.00

Note: Data at 6400 and 25600 ISO is unreliable because of implausible g% value.

We see the read noise starts at 5.16 e- then drops by about 1.9 e- at 200 ISO. Then it drops about another 0.5 e- to about 2.75 e- at 1600 ISO and above.

Compare this to the Canon 1DX table below:

Canon1DX

ISO: 100 200 400 800 1600 3200 6400 12800 25600 51200 102400 204800

m = 90101 46728 23720 11969 6296 3072 1524 738 332 163 104 52

n = 38.51 19.23 9.98 5.28 3.58 2.19 1.89 1.62 1.27 1.70 1.65 1.17

g% = 0.44 0.43 0.52 0.59 0.71 0.83 0.84 0.76 -0.00 -0.00 5.92 8.38

Note: Data above 12800 ISO is unreliable because of implausible g% value.

Here the read noise starts much higher at 38.5 e- at base ISO, then steadily drops to about 1.6 e- at 12800 ISO. So ISO gain does have a much greater effect on read noise for this Canon camera than for the Nikon D800(E).
Click to expand...
...I was basing my comments on sensorgen's figures for the D800:

http://www.sensorgen.info/NikonD800.html

which show a flat read noise response with regard to ISO, more or less.
 
Great Bustard said:
Great Bustard wrote:
Great Bustard said:
DSPographer wrote:

The D800(E) sensor is not ISOless. If it was, then we would see the DXOmark dynamic range plot drop by 1 Ev for each stop increase in ISO. In fact it drops only about 1/2 Ev from 100 to 200 ISO and drops 2.7 Ev instead of 3 Ev from 200 to 1600 ISO. Above 1600 ISO the D800(E) is ISOless. Another way to check for an ISOless sensor is to check the read noise in e- derived from the full SNR data. For the D800E that is the noise n in the following table (m is full well count and g% is PRNU):

NikonD800E

ISO: 100 200 400 800 1600 3200 6400 12800 25600

m = 54924 26293 13414 6722 3483 1776 833 438 237

n = 5.16 3.25 3.27 2.95 2.72 2.78 2.41 2.81 2.83

g% = 0.30 0.37 0.42 0.50 0.51 0.74 0.00 0.49 0.00

Note: Data at 6400 and 25600 ISO is unreliable because of implausible g% value.

We see the read noise starts at 5.16 e- then drops by about 1.9 e- at 200 ISO. Then it drops about another 0.5 e- to about 2.75 e- at 1600 ISO and above.

Compare this to the Canon 1DX table below:

Canon1DX

ISO: 100 200 400 800 1600 3200 6400 12800 25600 51200 102400 204800

m = 90101 46728 23720 11969 6296 3072 1524 738 332 163 104 52

n = 38.51 19.23 9.98 5.28 3.58 2.19 1.89 1.62 1.27 1.70 1.65 1.17

g% = 0.44 0.43 0.52 0.59 0.71 0.83 0.84 0.76 -0.00 -0.00 5.92 8.38

Note: Data above 12800 ISO is unreliable because of implausible g% value.

Here the read noise starts much higher at 38.5 e- at base ISO, then steadily drops to about 1.6 e- at 12800 ISO. So ISO gain does have a much greater effect on read noise for this Canon camera than for the Nikon D800(E).
Click to expand...
...I was basing my comments on sensorgen's figures for the D800:

http://www.sensorgen.info/NikonD800.html

which show a flat read noise response with regard to ISO, more or less.
Click to expand...
Yes, a bit strange the difference is that big at ISO 100 (5.16 vs. 2.7 e-), since they both are derived from the same DxO data.
 
Member said:
Ubilam wrote:

I see something that seems to be ignored here. The same camera can have more noise just a step or two up and then its clean again...ie: iso 100 can be as clean 320. 125 as bad as 400. Some high ISOs near to each other vary too more or less. Whats up with that?
Click to expand...
What are you looking at? Many camera JPEGs and/or external converters apply more noise reduction at higher ISOs, and/or in the shadows, so what you see isn't really what the camera is recording. Also, some cameras, like most non-1-series Canon DSLRs, use digital gain for 1/3-stop ISOs like 125, 160, 250, etc, and can have their post-gain, late stage noises boosted for a lower ISO than another ISO at which that noise can be attenuated, such that ISO 125 could have 2/3 stop more shadow noise than ISO 160, or even ISO 320. In fact, the dark frame noise on many Canon DSLRs is as high at 125 as it is at ISO 640 or even 1250. Banding noise is more easily visible when other noises are low, especially for images displayed very small. People have reported using ISO 640 on their Canon 5Dmk2 where they would normally have chosen base ISO (100) because even though the late stage noises are the same at both ISOs, the higher photon shot noise of 640 masks the banding that is visible at ISO 100.

Then, you have the issue of thresholds. Noise may be such a small part of the IQ at high exposures that doubling or halving it is imperceptible.
 
Last edited:
Steen Bay said:
Steen Bay wrote:
Steen Bay said:
Great Bustard wrote: ...I was basing my comments on sensorgen's figures for the D800: http://www.sensorgen.info/NikonD800.html which show a flat read noise response with regard to ISO, more or less.
Click to expand...
Yes, a bit strange the difference is that big at ISO 100 (5.16 vs. 2.7 e-), since they both are derived from the same DxO data.
Click to expand...
I prefer to use Bill Claff's read noise data for absolute comparisons because DxO's appears to be a milkshake of the four channels, possibly or possibly not scaled/balanced. Too many variables. Also it's difficult to read their SNR data in the deepest shadows because they do not display it with enough resolution (there is one stop between 0.01 and 0.02%) and I have a feeling that's one of the reasons for sometimes unreasonable sensorgen read noise values.

DxO data is of course fine when comparing relative data since the methodology is the same. I wish they gave us accesss to the individual channel data and told us how they actually compute SNR, though.
 
Last edited:
Jack Hogan said:
Jack Hogan wrote:
Jack Hogan said:
Steen Bay wrote:
Jack Hogan said:
Great Bustard wrote: ...I was basing my comments on sensorgen's figures for the D800: http://www.sensorgen.info/NikonD800.html which show a flat read noise response with regard to ISO, more or less.
Click to expand...
Yes, a bit strange the difference is that big at ISO 100 (5.16 vs. 2.7 e-), since they both are derived from the same DxO data.
Click to expand...
I prefer to use Bill Claff's read noise data for absolute comparisons because DxO's appears to be a milkshake of the four channels, possibly or possibly not scaled/balanced. Too many variables. Also it's difficult to read their SNR data in the deepest shadows because they do not display it with enough resolution (there is one stop between 0.01 and 0.02%) and I have a feeling that's one of the reasons for sometimes unreasonable sensorgen read noise values.

DxO data is of course fine when comparing relative data since the methodology is the same. I wish they gave us accesss to the individual channel data and told us how they actually compute SNR, though.
Click to expand...
Thanks, didn't know that Bill Claff had calculated the read noise in electrons. According to him the D800 read noise is app. 4.2 e- at ISO 100, and decreasing to app. 2.7 e- at ISO 1600.
 
Great Bustard said:
Great Bustard wrote:
Great Bustard said:
...I was basing my comments on sensorgen's figures for the D800:
Click to expand...
http://www.sensorgen.info/NikonD800.html

which show a flat read noise response with regard to ISO, more or less.
Click to expand...
There is something wrong with sensorgen's numbers. He is attempting to fit DXOmark's data, but somehow comes up with different values than I do. My fit to DXOmark's data is exact- which shows that they plotting a fit themselves instead of raw data. Since my fit to the DXOmark plot data is exact, I think sensorgen must be doing something wrong to get different values.
 
Jack Hogan said:
Jack Hogan wrote:
Jack Hogan said:
I prefer to use Bill Claff's read noise data for absolute comparisons because DxO's appears to be a milkshake of the four channels, possibly or possibly not scaled/balanced. Too many variables. Also it's difficult to read their SNR data in the deepest shadows because they do not display it with enough resolution (there is one stop between 0.01 and 0.02%) and I have a feeling that's one of the reasons for sometimes unreasonable sensorgen read noise values.
Click to expand...
Click to expand...
It doesn't matter how finely the DXOmark data is plotted when computing a fit. That is because their plot is just a mathematical curve generated from their estimates for read noise, photon gain, and PRNU for a sensor. As long as I use at least 3 plotted points when making a curve fit, I get the exact same answer for noise e-, full well e- and PRNU. (The match error residues in my overdetermined fits are consistent with double precision accuracy.
Jack Hogan said:
DxO data is of course fine when comparing relative data since the methodology is the same. I wish they gave us accesss to the individual channel data and told us how they actually compute SNR, though.
Click to expand...
I really wish they just gave as access to their raw data: but since they spent the effort to get it, they can choose what they want to release.
 
You must log in or register to reply here.

Keyboard shortcuts

f
Forum
w
Next unread
m
My threads
Mobile site
About
Editorial content
Cameras & Lenses
Community
All content, design, and layout are Copyright © 1998–2025 Digital Photography Review All Rights Reserved.
Reproduction in whole or part in any form or medium without specific written permission is prohibited.
When you use DPReview links to buy products, the site may earn a commission.
©GPS Media - Guides, Products, Services.
Back
Top