ISO vs DxO or DxO vs ISO

this might be covered to death before, too lazy to search

take same raw file, duplicate and change ISO to different values - one will be ISO 100 and one will be ISO 12800 - does DxO pay attention to a tag vs purely raw data when it demosaick and DPXD ? it does , no - better to see in full resolution = ISO DXO — Postimages


repeated with applying NO CORRECTIONS preset followed by enabling default DPXD

full size = ISO DXO one more — Postimages


Thank you.

the point ( requires additional testing of course ) is - if DxO PL (or PoorRaw) is your raw converter, do you want to use ISO as camera suggests -OR- you want to use the lowest ISO where S/N gain is good ( like 2nd base ISO in dual gain sensors ) and push … given that apparently (again - more tests needed ) DxO behind the scenes treats the same raw data differently and higher ISO value in a tag leads to details loss ( same DPXD parameters used in UI ) - may be you want to use a lower nominal ISO and push brightness in converter ( that though might have other adverse effects )…

A7M4 raw file @ ISO400 and underexposed

duplicated 4 times and put ISO tag value as ISO 400, ISO1600, ISO6400, ISO51200

applied no corrections preset and brighten the image just with a tone curve tool ( same for all 4 raws )

full size = DxO 1 — Postimages


nothing drastic, except what DxO does with the same raw date when it sees ISO51200 written in the file

(GIF is animated as : ISO400 - 5sec delay then cycle through ISO1600-ISO6400-ISO51200 with 2 sec for each)

I was going to repeat the same with DPXD when I found that @ ISO51200 DPXD simply goes nuts

see screenshot - pay attention to preview window in NR tool in the left side docking place

full size = Dx O DPXD dies at ISO51200 — Postimages

so humans still stand a chance to kill AI when necessary :smile:

And Adobe ? … you knew !!!

full size screenshot = ADOBE with flying colours — Postimages

it seems no chance for humans if AI is done by Adobe… all hail DxO AI !!!


A7M4 raw file @ ISO400 and underexposed

duplicated 4 times and put ISO tag value as ISO 400, ISO1600, ISO6400, ISO51200

applied no corrections preset and brighten the image just with a tone curve tool ( same for all 4 raws ) and then enabled default DPXD processing … remember all the same raw data inside… just different values for ISO exif tag manually altered

full size = DxO 2 — Postimages


(GIF is animated as : ISO400 - 5sec delay then cycle through ISO1600-ISO6400-ISO51200 with 2 sec for each)

and now the same with ACR and its default Enhanced NR settings

full size = ACR Enhanced NR Default — Postimages


(GIF is animated as : ISO400 - 5sec delay then cycle through ISO1600-ISO6400-ISO51200 with 2 sec for each)

PS: do u see any material changes in Adobe’s processing ?!

This repeted for a photo of an organic motif - say forrest, architecture etc would be a good next step as the ML based AI was trained on those kind of photos and not text.

true indeed that bad stuff difficult to find in the forest … and Adobe trained how ?

more so - we see that stuff goes bad not based on raw data, but based on what ? some tag in exif ? that you explain how ? may be a better way is to pay attention to the image data, not making assumption what to do just based on the tag …

in any case it is a food for thought… further testing shall be done

I would think it makes sense to process the data differently based on the recorded ISO, as having prior knowledge of noise characteristics at different ISO settings would seem, to me, to be a good input.

Imagine, for example, that your sensor is demonstrated to produce more noise in the red pixels when ISO goes above 6400. Possibly an unlikely scenario, but please bear with me.

If DxO “know” your image was taken at ISO 12800, then they can be more aggressive with (or at least more cognisant of) removal of what appear to be red noise signals.

Now that is purely a thought experiment as to why ISO level (which after all can only ever be ‘just a tag’) may affect the output. However, I am reminded of a discussion in the Pentax forums I frequent, where it was claimed (and I am inclined to believe) that the modern Pentax sensors (made by Sony, I believe) are effectively ISO-invariant.

The discussion was centred on the technique of always shooting at base ISO and pushing exposure in post, for the benefits that can bring. For the purposes of this discussion, I figure the claim that Pentax (Sony) sensors are ISO-invariant implies that some sensors are not. In which case my thought experiment above makes some sense.

of course it makes sense to pay attention to a nominal ISO in terms of what we have in deep shadows - that is what analog gain and/or dual gain affects … however it seems there is too much attention paid to a tag value to the detriment of actual data ( which seems Adobe avoids ) … and in most ( if not ALL ) modern cameras ( recent Pentax cameras with their big mandatory NR for raw data are sore exceptions ) that difference say past ISO800 seems to be non material ( of course there might be a pattern noise for example that testing sites like photonstophotos from BClaff do not test for , as it happens well above individual sensel level to see ) …

shall try more testing when time permits - say the same scene, “A” mode with fixed ISO, changing nominal ISO values across the range, then let camera meter exposure to set exposure time and dial (A) positive EC ( just barely avoiding clipping in raw ) and (B) negative EC (say to have 3-4 stops “underexposure”)… then repeat to see how DPXD works… ultimately whatever happens one needs to know if it makes sense to venture in to say past “ISO6400” nominal ISO values when shooting if the shot intended for DxO DPXD NR or stay put ( same exposure, but just not crossing “ISO6400”… 6400 is used just as an example )

It would be interesting to see an iso 400 image and an iso 12800 image of the same scene. With their original tags and reversed tags. Then you could see if that tag has any advantages or disadvantages.

It’s a neat experiment - a way of looking under the covers at how DxO’s NR works. But testing ISO invariance is where this will really get interesting, IMO. That has real value for our photography and how we use PhotoLab.

I was surprised too that ISO value in the exifdata didn’t change denoise value’s in the tool it stayed 40 when magic want was active. Which would imply “automatic”

Too long ago to remember the details. I think the original thread was even older.
I need to go to work so can’t dig deeper at this time.
Something about they remove the noisyparts before demosiacing and therefore before lighting/brighting is applied which would also be applied on the noise particles.
That’s why ISO value isn’t effecting automated deniosestrenght…
I believe that was dxo’s explaination.
From that moment i just made partial presets in my camera’s iso sections.
So when i have a lot of high iso shots i just apply that.

I don’t understand what is meant in this thread. I did create 6 copies of an image and changed the iso from 100 to 51200. No differences. And when I think about it, this is as it should.


1 Like

but it seems it does affect how NR works ( same UI slider value )… because I only changed ISO value post factum - nothing else in raw data was different…

if I follow “exposure triangle” and give less and less actual exposure to sensor ( but increasing nominal ISO values ) then as expected DxO NR starts to meltdown at high ISO values ( because actual exposure is less and less )… so this is expected… but with the same data (as it is the same raw) just altering ISO values to have NR meltdown was not expected.

PS: and going to lower ISO value in the tag visibly reduce the amount of NR applied to the same raw file all other things equal …

So, i haven’t test it myself, with the same exposure and same imagedata except iso value in exifdata dxo’s NR algorithm is tricked by your alter and react accordingly to the iso value. That’s different then they wrote earlier.
So the strenght slider isn’t connected with the iso value in exif butunder the hood they let there “AI” known which isovalue is used so it changes approach.

DxO might have different algorithms or approaches for different kinds of raw files depending on sensor data specifications and manufacturer etc.