Originally posted by gatorguy I don't think that bolded part is entirely accurate. Have you tried DxO which presumably is tuned to the specific camera/processor/lens? AFAIK Adobe does the same, having profiles for specific camera/processor combinations in Camera Raw. Capture One definitely does too. A cheaper software such as Luminar may simply average things out. No idea about them.
It was a bit of a blanket statement. It depends on the extent to which the NR algorithm is actually tuned/profiled for the specific sensor. For instance I once dug into the papers behind the Darktable NR algorithm (as I went through the creation of a profile for a camera missing from the database). That specific solution was based on an analytical model whose parameters are tuned to each camera, starting from pictures taken to a very specific target. Such model was grounded on assumptions about the statistical properties of sensor noise, which are expected to hold true for "most sensors". If there is aggressive noise reduction baked into the raw files that makes a sensor violate such assumptions, the algorithm might (on paper) perform poorly even with the best available estimate for the parameters. Regarding more advanced algorithms like DxO DeepPrime, that incorporate deep learning models (supposedly) at a certain stage of the pipeline, it depends on how the per-camera customization actually takes place. Is the model (I speculate, some sort of convolutional neural network) fully, or partially, retrained on a new dataset for each camera ? Is it fixed, completely general, and just some additional parameters unrelated to the DL part of the pipeline are fine-tuned to the cameras ? We will likely never find out the answers to these questions. The only place where some truly ad-hoc processing is guaranteed to take place is the camera, and this is why I fully agree with this statament:
Originally posted by gatorguy Doing so in-camera, carefully matched during development with the actual chosen processor traits and shortfalls (there's always some), makes more sense to me than NR defaults in a third party RAW importer that isn't tuned to the camera specifics.
I have no a-priori prejudice against this technique, if the NR software is properly tuned, then aggressive in-camera noise reduction could the best performing solution even from a technical point of view. What I was implying is rather the fact that those DR measurements are to be taken with a grain of salt when comparing sensors where such aggressive NR is implemented and with traditional designs. Not because one solution is inherently "wrong", but because it's an apples vs. oranges comparison. Noise reduction alters that DR "measurement" in such a way that the comparison is unfair or I'd better say unreliable, inconclusive.
I have DxO Photolab 5; actually I went again through some studio test samples comparing K1 vs K1mk2 using this handy tool
GitHub - rdaforno/cropcomparator: script to compare crops of different photos, runs offline in a browser at high iso such as 12800, and different NR settings. I'm still convinced that the K1 original, with "all" NR duty handled by DxO allows for the best compromise between detail retention and reduction of grain in uniformly colored areas. By a considerable margin ? No, it's far from catastrophic. It depends on the portion of the scene, on some crops there is almost no difference. On some very specific areas the detail loss is apparent. This is not unexpected as
any NR algorithm has performances that vary depending on the specific scene.
Is it a complete show stopper ? Should users be advised against buying the mark2 ? Absolutely not. But is there any region of the test frames I've looked at where the mark 2 delivers better high-iso performance ? I don't see it. At best it equals the Mk1 model. It's clearly a solution developed to benefit Jpeg shooting. If you have access to best-in-class NR software, there's no benefit that I can see. Do I believe that on a stills-oriented full frame camera, clearly geared towards class-leading image quality there should be an option to turn this feature off? Yes, firmly.