New parts of a (Subject) mask appear after/because denoising Apple ProRAW DNG (iPhone 15 Pro)

[germans67014243]

Just updated to LRC 14.0 and while testing Apple DNG denoise noticed it adds new objects to my Subject mask (after/because AI mask update?). In any case, this is an issue.

First photo is the original mask, yellow marks in the second photo - "new" parts of that mask. What's going on?

[Rikk Flohr: Photography]

Can you export this file as DNG with the masking intact and share it with us for our engineers to review?

[germans67014243]

https://we.tl/t-Bmxc4zu6ES

[Simon S Chen · Accepted solution]

This is expected because the AI denoise is processed upstream to the ML masking. If you apply denoise first, that would affect the what the ML masking model (ie select subject) sees. This is as designed.

[Simon S Chen]

After applying the denoise, it makes the select subject model see the underlying structure more clearly.

[germans67014243]

I always apply denoise at the end of my workflow. I mean what's the point of applying denoise at the start of the editing when more noise can be introduced by subsequent operations (+shadows, dehaze, etc.), and at that point I can not apply denoise again? Or can I now, in LRC 14? Even so, that would mean additional DNG file and disk space waste.

[Conrad_C]

Applying denoise late in the workflow is the opposite of what Adobe recommends in their tech article Denoise Demystified (scroll down to Best Practices), which was written by one of the lead engineers. He also talks about possibly having to revisit edits such as Shadows, Clarity, or Dehaze if you apply AI Denoise later — another argument to do it sooner.

I am far from qualified to be an engineer so maybe an engineer can correct me if I am wrong in the explanation below, but my understanding is that “more noise can be introduced by subsequent operations (+shadows, dehaze, etc.)” is not technically correct.

As I understand it, a raw image has a fixed amount of noise based on the sensor and the amount of gain applied. Although a lot of people say that various adjustments in raw software can “increase noise” or “introduce noise,” that’s probably a misinterpretation or mis-stating of what’s happening. What really happens is that various adjustments can make more visible the same noise that was always there. It’s like in audio engineering where the noise floor is too high: Your audio adjustments are not adding noise, just amplifying the noise that was in the recording from the beginning.

So it’s better to do AI Denoise first, setting the value to a good balance between too low (distracting noise) and too high (plasticky), giving all the tools that improved, lower noise floor to start from. If subsequent edits make the noise more visible, one thing I do is apply Noise Reduction through a mask to restrict it to where the noise was amplified. Sometimes, this can be done in the same mask that introduced the noise. For example, lightening the shadows can make existing noise more visible (because the darkest bits of a digital image are always the noisiest), so if that was done through a mask, I might increase the Noise Reduction value in that same mask to compensate.

If it’s decided that the original AI Denoise value was too low, I will re-do AI Denoise, sync the edits from the old version to the new one, and look closely for any edits that need adjusting for the new noise level.

[germans67014243]

This looks like a much more complicated workflow, expecially in the last paragraph.

[Conrad_C]

This looks like a much more complicated workflow, expecially in the last paragraph.

By @germans67014243

Actually I think it’s simpler, because once you understand that it’s all about the starting level of noise (because none is added later), the only thing you have to do is get the noise level right at the first step, AI Denoise. And that can be previewed before applying. So I almost never have to re-do AI Denoise.

And where that last paragraph really came from was not re-doing AI Denoise, but to handle the situation where, for example, I edited an image 3 years ago before AI Denoise existed, and I want to make a much cleaner version of it now that AI Denoise is available. The technique in the last paragraph means that I can run AI Denoise on the same original and not have to re-do the entire edit that took me a long time to work out, since syncing only takes a few seconds. So I see the last paragraph as a technique that has saved me hours of time. In the new version, I only have to roll back edits that tried to compensate for the noise in the older version.

One extremely important reason to denoise early is because of something Photoshop/Camera Raw expert Jeff Schewe pointed out many years ago: Noise and sharpening are interrelated, so if you want to evaluate and correct sharpening properly, you should have already reduced noise by that point. Not later in the workflow.

[germans67014243]

That's true, I should correct myself - I apply denoise as the step before the last one, which is sharpening indeed. As for "...none is added later" - I need to check this myself.

[Simon S Chen]

This is a valid concern that could be addressed in the future.

[Simon S Chen]

See the https://helpx.adobe.com/camera-raw/using/whats-new/2025.html. You can use Denoise, Raw Details, and Super Resolution non-destructively in the Detail panel without creating a new DNG file. It is a technology preview in the Adobe Camera Raw plug-in only initially.