Canon EOS R1: The Best High ISO Performance Yet from a Canon sensor

AlanF said:

Those tests are not done on RAW output, they are done on jpegs so they are not answering your question.

Click to expand...

Visual comparisons like the DPR comparison tool or the digital picture one are presented as RBG render (compressed to jpeg). DxOMark and photonstophotos work with raw files. Dynamic range requires knowing the noise floor and you need raw files for that.

AlanF said:

I've tried most of the RAW converters and I find DxO extracts the most detail and doesn't add detail as you imply - it's not Topaz AI.

Click to expand...

More noise (technically - lower signal to noise ratio) means less information. That information cannot be recovered. It can only be guessed and/or generated. When you see noise in an image and then no noise after processing, that means fake detail was added.

DxO actually uses AI afaik, but that doesn't matter - the detail either gets generated by AI or algorithmically 'guessed'.

Quarkcharmed said:

Visual comparisons like the DPR comparison tool or the digital picture one are presented as RBG render (compressed to jpeg). DxOMark and photonstophotos work with raw files. Dynamic range requires knowing the noise floor and you need raw files for that.

More noise (technically - lower signal to noise ratio) means less information. That information cannot be recovered. It can only be guessed and/or generated. When you see noise in an image and then no noise after processing, that means fake detail was added.

DxO actually uses AI afaik, but that doesn't matter - the detail either gets generated by AI or algorithmically 'guessed'.

Click to expand...

My demands are pretty high for image production because I usually crop like mad and use high isos for high shutter speeds. I have DPP4, LR. PS, TopazAI and DxO PL and have compared them directly as RAW converters on subjects with high levels of detail under different light levels to see how much detail and noise and whether there appear to be artefacts. I've also occasionally stacked a series of noisy images and used PS statistical functions to reduce noise. For uncropped images under normal conditions, all can give excellent results. At the extremes at which I work, DPP4 is pretty hopeless at noise reduction. Topaz AI and sharpen are prone to introduce artefacts. PS has improved in recent years at noise reduction, and DxO Photolab produces the cleanest images with the most detail without my seeing any artefacts. What do you see in your comparisons?

AlanF said:

My demands are pretty high for image production because I usually crop like mad and use high isos for high shutter speeds. I have DPP4, LR. PS, TopazAI and DxO PL and have compared them directly as RAW converters on subjects with high levels of detail under different light levels to see how much detail and noise and whether there appear to be artefacts. I've also occasionally stacked a series of noisy images and used PS statistical functions to reduce noise. For uncropped images under normal conditions, all can give excellent results. At the extremes at which I work, DPP4 is pretty hopeless at noise reduction. Topaz AI and sharpen are prone to introduce artefacts. PS has improved in recent years at noise reduction, and DxO Photolab produces the cleanest images with the most detail without my seeing any artefacts. What do you see in your comparisons?

Click to expand...

I have also found that DXO Photolabs makes very clean images with very few (if any) artifacts. Similar to your experience I have noticed that Topaz Photo AI can introduce artifacts so one needs to be careful. Overall, the two programs I use for RAW development are DXO Photolabs and Topaz Photo AI.

AlanF said:

My demands are pretty high for image production because I usually crop like mad and use high isos for high shutter speeds. I have DPP4, LR. PS, TopazAI and DxO PL and have compared them directly as RAW converters on subjects with high levels of detail under different light levels to see how much detail and noise and whether there appear to be artefacts. I've also occasionally stacked a series of noisy images and used PS statistical functions to reduce noise. For uncropped images under normal conditions, all can give excellent results. At the extremes at which I work, DPP4 is pretty hopeless at noise reduction. Topaz AI and sharpen are prone to introduce artefacts. PS has improved in recent years at noise reduction, and DxO Photolab produces the cleanest images with the most detail without my seeing any artefacts. What do you see in your comparisons?

Click to expand...

The topic is about the 'sensor performance'.
What you describe above is the comparison of software, not sensor performance.

If DxO wins for you - great, but if the real detail is lost, there's simply no way to recover it. The laws of nature don't allow that unfortunately. What you see after applying DxO is plausible but fake detail - either guessed or generated.

john1970 said:

I have also found that DXO Photolabs makes very clean images with very few (if any) artifacts. Similar to your experience I have noticed that Topaz Photo AI can introduce artifacts so one needs to be careful. Overall, the two programs I use for RAW development are DXO Photolabs and Topaz Photo AI.

Click to expand...

That's fantastic, but it has nothing to do with the sensor performance measurements.

Let's say you have images from sensors A and B, and the sensor B has a slightly lower dynamic range and therefore a bit higher visible noise.

But after DxO you see the same level of detail and quality in A and B cases. All that means is either you didn't notice the difference, or DxO applied a bit more of noise reduction and generated a little bit more of fake detail for the sensor B.

Quarkcharmed said:

The topic is about the 'sensor performance'.
What you describe above is the comparison of software, not sensor performance.

If DxO wins for you - great, but if the real detail is lost, there's simply no way to recover it. The laws of nature don't allow that unfortunately. What you see after applying DxO is plausible but fake detail - either guessed or generated.

Click to expand...

Would you please answer the simple question I asked that you have ignored, which I will repeat more explicitly: have you done a series of experiments varying the exposure of an image with high detail that gets progressively more noisy and then looked at the detail recovered by different software? Yes or no?

AlanF said:

Would you please answer the simple question I asked that you have ignored, which I will repeat more explicitly: have you done a series of experiments varying the exposure of an image with high detail that gets progressively more noisy and then looked at the detail recovered by different software? Yes or no?

Click to expand...

Yes.
But I was comparing denoise software for my purposes, not sensors.

Quarkcharmed said:

Yes.
But I was comparing denoise software for my purposes, not sensors.

Click to expand...

So, did you find DxO PL was adding false detail?

AlanF said:

So, did you find DxO PL was adding false detail?

Click to expand...

They all add fake detail one way or another. I'm not saying 'false' btw. That would imply the added detail were completely wrong. Guessed/generated detail may be complete off or may be close to would-be real detail.

Since Adobe introduced AI denoise, I use it for noisy (high ISO) images.

Quarkcharmed said:

They all add fake detail one way or another. I'm not saying 'false' btw. That would imply the added detail were completely wrong. Guessed/generated detail may be complete off or may be close to would-be real detail.

Since Adobe introduced AI denoise, I use it for noisy (high ISO) images.

Click to expand...

Ok, you haven’t done the experiments to find out whether in practice DxO denoising adds fake detail. Max Planck famously said: “Experiment is the only means of knowledge at our disposal. Everything else is poetry, imagination.”

AlanF said:

Ok, you haven’t done the experiments to find out whether in practice DxO denoising adds fake detail. Max Planck famously said: “Experiment is the only means of knowledge at our disposal. Everything else is poetry, imagination.”

Click to expand...

You might as well conduct an experiment with apples. Add 2 and 2 apples and check if there's 4 of them now. Max Planck implied the study of the unknown laws of the universe, not a computer program like DxO or Topaz denoise, created by humans.

DxO denoise isn't a black box. DxO developers know it better, but we already know how it works in principle. You don't need to do experiments in order to conclude there's 'fake' detail. It's just a given that this software generates/guesses the detail because it's how it works. It's not matter of experimentation, it's already a mathematical truth.

I understand DxO works great for you - it's a great tool indeed, but you're missing my original point (see above in the thread) - sensor performance comparisons are not done on heavily processed images with generated/guessed detail. This rules out software like DxO, Topaz etc. and any other noise reduction software with or without AI.

Ok if you really need a visualisation, here it is. The images from DPR scene comparison - raw images from the R5II at ISO 100 and 25600. They illustrate the loss of detail and what I call 'fake' detail. It's your beloved DxO.

Unprosessed images. R5II, ISO 100 vs ISO 25600. Obviously, at ISO 100 we see more detail, the text on the thread rolls is readable. Not so much at ISO 25600



Imaged processed in DxO PureRAW 4. R5II, ISO 100 vs ISO 25600. After processing, the text is still readable at ISO 100 and still unreadable at ISO 25600. It's kinda sharp and looks plausible at the first glance but doesn't make any sense if you try to read it. Thread loops are not separated and almost completely lost. That's the generated detail.



So above you can see an illustration of a very simple fact: DxO cannot actually recover information that was lost. You can see the difference between 'more' information at ISO 100 and 'much less' information at ISO 25600.

However, if you try and measure the visible noise (as standard deviation) in the images on the bottom, it'll be roughly the same. Not only we lose information in the images, we also lose any usable information about the sensor's performance.

Any of you having issue with False Colour on your R1? When I have false colour engaged, and try to record, the screen freezes and the camera doesn't display recordings properly until I turn off and on. Otherwise, zero issues. I've tried different cards, lenses, resetting the camera, etc. Trying to decide if I should ship back and exchange or hope for a firmware update to resolve it.

Thanks for any help! If any of you have an R1 and could test your false colour for me (while recording) that would be super helpful!

Quarkcharmed said:

You might as well conduct an experiment with apples. Add 2 and 2 apples and check if there's 4 of them now. Max Planck implied the study of the unknown laws of the universe, not a computer program like DxO or Topaz denoise, created by humans.

DxO denoise isn't a black box. DxO developers know it better, but we already know how it works in principle. You don't need to do experiments in order to conclude there's 'fake' detail. It's just a given that this software generates/guesses the detail because it's how it works. It's not matter of experimentation, it's already a mathematical truth.

I understand DxO works great for you - it's a great tool indeed, but you're missing my original point (see above in the thread) - sensor performance comparisons are not done on heavily processed images with generated/guessed detail. This rules out software like DxO, Topaz etc. and any other noise reduction software with or without AI.

Ok if you really need a visualisation, here it is. The images from DPR scene comparison - raw images from the R5II at ISO 100 and 25600. They illustrate the loss of detail and what I call 'fake' detail. It's your beloved DxO.

Unprosessed images. R5II, ISO 100 vs ISO 25600. Obviously, at ISO 100 we see more detail, the text on the thread rolls is readable. Not so much at ISO 25600
View attachment 221160


Imaged processed in DxO PureRAW 4. R5II, ISO 100 vs ISO 25600. After processing, the text is still readable at ISO 100 and still unreadable at ISO 25600. It's kinda sharp and looks plausible at the first glance but doesn't make any sense if you try to read it. Thread loops are not separated and almost completely lost. That's the generated detail.

View attachment 221159

So above you can see an illustration of a very simple fact: DxO cannot actually recover information that was lost. You can see the difference between 'more' information at ISO 100 and 'much less' information at ISO 25600.

However, if you try and measure the visible noise (as standard deviation) in the images on the bottom, it'll be roughly the same. Not only we lose information in the images, we also lose any usable information about the sensor's performance.

Click to expand...

You have misrepresented what I have written. I have never claimed at any point that the software restores all the detail at any iso. Indeed no one claims that denoising software does this - it just makes the most of what data can be recovered. What I have written is that DxO extracts the most detail of the software I have used without giving noticeable artefacts. Of course as you go to higher and higher iso, you lose more detail, and what the software does is to reveal more of the underlying detail that is not lost and is still recoverable. No one in their right mind would take an image at iso 25600 and expect not to lose resolution compared with one at iso 100. An appropriate series of experiments would be looking at objects that are not monochrome and monotonous but have proper details and contrast on them, and cover a scientifically logical range of values like iso 100, 200, 400, 800, 1600, 3200, 6400, 12,800 and 25,600 so you can see how detail is progressively lost with and without software intervention in an iso range that would cover your normal operating range. (My most frequently used is iso 800, going up to higher when necessary.) My usual subjects are birds with exquisite details in their plumage, and I see loss of detail with increasing iso with all software but I don't see new "fake" details appearing with DxO.

Regarding experiments and knowledge, as expounded most clearly by Karl Popper, it's the tenet that all theories must be falsfiable, ie must be able to be subject to experimental testing to see whether they fit the facts - not just the fundamental laws of the universe but everything. Without experimental testing, a proposal ("theory") is just a hypothesis.

AlanF said:

You have misrepresented what I have written. I have never claimed at any point that the software restores all the detail at any iso.

Click to expand...

I wasn't interpreting your messages as claims that the software restores 'all' details. But you were arguing against the fact there's fake detail - fake in the sense I explained above.

AlanF said:

Indeed no one claims that denoising software does this - it just makes the most of what data can be recovered. What I have written is that DxO extracts the most detail of the software I have used without giving noticeable artefacts.

Click to expand...

I provided examples with artifacts above. The recovery as such is very limited. 'Recovery' would be something you didn't see in the original image but do see in the processed one. Denoise software can reduce colour noise which improves the image in terms of recognisability, because the colour noise really hides the detail that can be 'recovered'. But after reducing the colour noise, there's very little you can actually recover with any software.

DxO makes noisy images visually more pleasing and sharp - but by generating or guessing the fake detail.

Check the samples above at ISO 25600, before and after denoising. What was unreadable stayed unreadable, mostly. And what was readable stayed readable. But on top of that, DxO generated a lot of fake detail to make the image visually sharper and more detailed.

AlanF said:

Of course as you go to higher and higher iso, you lose more detail, and what the software does is to reveal more of the underlying detail that is not lost and is still recoverable.

Click to expand...

What you call 'recovery' is either something you could see already without DxO, or AI-generated detail.

AlanF said:

An appropriate series of experiments would be looking at objects that are not monochrome and monotonous but have proper details and contrast on them,

Click to expand...

That sounds like you want to introduce a selection bias. What do you mean by 'proper' detail and contrast? You think the thread rolls in the samples above didn't have enough contrast? Enough for what?

The threads were perfectly discernable at ISO 100, so I'd say it was proper detail and contrast. Then at ISO 25600 the information that allowed to show the thread separation was lost and we can see, very clearly, that DxO failed to recover it. Instead it created a lot of fake pixels. You can also see fake pixels in the text on the rolls - the letters are distorted.

AlanF said:

and cover a scientifically logical range of values like iso 100, 200, 400, 800, 1600, 3200, 6400, 12,800 and 25,600 so you can see how detail is progressively lost with and without software intervention in an iso range that would cover your normal operating range.

Click to expand...

As I said, in the example above, ISO 100 shows the image with "a lot" of information near maximum that the R5II can get, and ISO 25600 shows the image with less information - something like 6 bits per pixel less in raw.

AlanF said:

Regarding experiments and knowledge, as expounded most clearly by Karl Popper, it's the tenet that all theories must be falsfiable, ie must be able to be subject to experimental testing to see whether they fit the facts - not just the fundamental laws of the universe but everything. Without experimental testing, a proposal ("theory") is just a hypothesis.

Click to expand...

Again we're not talking about hidden laws of the universe. We do know how this software works, there's no need for theories.

Quarkcharmed said:

Again we're not talking about hidden laws of the universe. We do know how this software works, there's no need for theories.

Click to expand...

I do know a little about the subject - my friends will spot me in the photo here - and I do not know what is going on in the neural networks in this software. https://www.chalmers.se/en/current/...ome-together-to-shape-the-future-of-research/

AlanF said:

I do know a little about the subject - my friends will spot me in the photo here - and I do not know what is going on in the neural networks in this software. https://www.chalmers.se/en/current/...ome-together-to-shape-the-future-of-research/

Click to expand...

Then I don't understand why you're saying there's no generated detail in DxO-processed images.

Neural networks do generate detail. AI-processed denoised images actually lose some of the original information, because the process is irreversible - it's not possible to restore the original image out of AI-processed one. The information that was lost is replaced by the generated detail. And that's visualised in the samples above.

Just a quick note: Thank you to those of you who have supported these pre-sets. The feedback I have received has been nothing short of exemplary. Greatly appreciated - thank you.