Ysarex wrote:
That photo is something of a parlor trick and doesn't really represent what most people can expect from that high an ISO. That photo comes from one of my classes. I took it specifically for my students as an illustration. I get a new crop of students every semester and they come already screwed-up by Youtube nonsense. Bottom line they believe that ISO causes noise. They're related no doubt but ISO or a high ISO is not the cause of the noise. The cause of the noise is the reduced exposure that typically goes with high ISO. That photo was in fact taken with the camera set to ISO 12,800 -- the EXIF data is intact and correct. But the kicker is that I applied a lot more exposure than would be expected at that ISO. With more exposure I got a lot less noise.
It's a parlor trick because if I could really add that additional exposure I really didn't need that high an ISO. I do it for my class to make the point: the noise isn't coming from what ISO does it's coming from the reduced exposure. I use the photo to illustrate that the belief that ISO causes noise is faulty.
Bottom line if you're really in a situation where you can't add more exposure and have to raise the ISO that high you're going to get a noisier photo than I showed here. But it's worth it here as well to make the same point. It's not ISO that's causing the noise.
That photo is something of a parlor trick and does... (
show quote)
Nerdy explanation of the ISO/noise relationship (without going too far into the gritty details):
Your sensor knows nothing about your ISO. It takes photos and converts them to electrons, which it stores. The number of photons depends on the intensity of the light and the time of the exposure. If you have "the right exposure" the brightest parts of your subject will produce enough electrons to max out the storage capacity of that particular pixel in the sensor. If you are overexposed, you have too many electrons so some of them will be thrown away. If you are underexposed, you will not reach the maximum capacity of the sensor.
So after the exposure your sensor has a bunch of areas with electrons stored in them. What now? You have to read out the sensor and convert the number of electrons at each pixel into a digital numeric value. You do that by taking the electrons and converting them to a voltage, which you then convert into a number. For a 12 bit sensor, the largest number you can get is 2^12, or 4096. For a 14 bit sensor, 2^14=16384.
At this point we get into the realm of statistics. For integers, the uncertainty of any counted number depends on Poisson statistics, but when the number is large enough, Poisson statistics merge into Gaussian statistics (usually around numbers greater than 10). For Gaussian statistics, the uncertainty in a given number is given by the square root of that number, generally called σ (sigma). Herein lies noise. If you have an area of uniform brightness in your image, with, say, a brightness that is converted into the number 1024, the square root of that is 32. The actual number you will get will average 1024, but can vary about that value. The variability is the noise.The variation of the number will fall within a Gaussian curve, so that ~68% of your numbers will be within σ of 1024, ~95% of your numbers will be within 2σ of 1024, ~99.7% will be within 3σ of 1024. So although your image has an area of uniform brightness in real life, your camera will show it with some brightness variations. For this example, σ is about 3% of the brightness. If your uniform brightness area is not so bright and only gives you 128 electrons instead of 1024, σ will be 16. So now σ is about 6% of the brightness, and you will see more variability, i.e. more noise. So noise is going to depend on the number of electrons your sensor is storing.
Now we have to consider what ISO does.
Raising the ISO basically multiplies the number that is derived when you read out the electrons from your sensor. It might just use an analog approach, in which the voltage derived from the number of electrons is raised by changing the amplification, or a digital approach in which the number derived from the number of electrons is just multiplied by a constant. Neither of these approaches will change the relationship between the final number and the variability of that number from the initial statistics.
The advantage of raising the ISO is that you can use shorter exposures to get what looks like good exposures in terms of numbers coming from the sensor readout. That means you can use a higher shutter speed to reduce motion blur, or a smaller aperture to increase depth of field. Modern cameras probably apply image processing software to reduce the noise so it will not look so bad, but that cannot eliminate the noise.
There are other sources of noise, not considered here, but they are small in normal conditions.