Thick red/blue banding in brightened shadows of Denoised DNGs

[PvlKnn]

I noticed this a few days ago but didn't realize the scale of the issue until I played around with the Exposure values and tested them to the limit.

I've attached two screenshots of the Develop module. The control image is an RAF photo with Exposure boosted to +5. The experiment is a DNG derived from the RAF, with a Denoise amount of 30. The effect manifests at all amounts from 1 to 100. However, the banding is not present in the Enhance Preview window. I have also attached the original Fujifilm RAF from my X-H2s camera (lossless compressed) and an affected DNG with Denoise.

I can't detect this phenomenon in NR conversions before LrC 13.3. 

The significance should be self-evident. The banding artifacts appear in images with lifted shadows or underexposed photos with boosted brightness.

Lightroom Classic version: 13.3.1 [ 202405311538-6d7c0308 ]
License: Creative Cloud
Language setting: en-CA
Operating system: Mac OS 14
Version: 14.5.0 [23F79]
Application architecture: arm64
Logical processor count: 12
Processor speed: NA
SqLite Version: 3.36.0
Power Source: Plugged In, 80%
Built-in memory: 32,768.0 MB
Dedicated GPU memory used by Lightroom: 5,764.7MB / 21,845.3MB (26%)
Real memory available to Lightroom: 32,768.0 MB
Real memory used by Lightroom: 3,716.8 MB (11.3%)
Virtual memory used by Lightroom: 424,322.1 MB
Memory cache size: 1,238.0MB
Internal Camera Raw version: 16.3.1 [ 1889 ]
Maximum thread count used by Camera Raw: 5
Camera Raw SIMD optimization: SSE2
Camera Raw virtual memory: 2281MB / 16383MB (13%)
Camera Raw real memory: 2285MB / 32768MB (6%)

[johnrellis]

The original raw didn't make it as an attachment. Upload it to Dropbox, Google Drive, or similar and post the sharing link here. Adobe likely won't pay attention without being able to reproduce the issue.

[Ian Lyons]

Your System Info indicates that you're using an Apple Silicon based Mac. LrC 13.3 was the first version to use the Apple Neural Engine, which is feature of Apple Silicon. This 'may' explain why the issue wasn't present in earlier versions. I'm not aware of any other changes to Denoise. 

Please, upload the original raw as @johnrellis requested. 

[PvlKnn]

Sorry for missing the raw files that were omitted from the attachment. You can download the original Fujifilm RAF and corresponding DNG (Denoise amount 30) from my OneDrive: 

https://1drv.ms/f/s!AjAOXAIjC1XCgaMEHGfxMYIMFnHMnw?e=lzb5o4

[Ian Lyons]

I downloaded your two files. and imported them both into a copy of LrC 13.2 and 13.3.1

When I ran denoise on the RAF file in LrC 13.2 I get a large purple block in the darker regions of the image. However, when I denoise the RAF file in 13.3.1 I get a grid of large purple and orange squares in the darker regions. Therefore, since both versions are producing the coloured blocks, albeit of different size and pattern, my original theory of the Apple Neural Engine being a possible cause is in doubt.

As a further test, I increased the exposure of the RAF file by 5 stops plus Shadows of 100%. After doing so, the darker areas have a heavy yellowish tinge.  At this point I'm inclined to say that rather than a bug,  the results obtained from denoise are being heavily influened by the underlying colour tinged areas in the original RAF file. 

Maybe @johnrellis can run a test using Windows to see if he produces similar results to mine.

I've attached a screenshot of denoised version from LrC 13.2 and 13.3.1.

[PvlKnn]

We can call it a bug or a result of how Denoise interprets the underlined colour in the deepest shadows, but the fact remains: in the deepest shadows, Denoise is generating colour patterns that are not present in the source raw file. It's not even doing it consistently across different computers, to boot. 

To ensure this wasn't caused by some aspect of the lights in the scene (doubtful), I Denoised and pushed a photo from months ago of a different location. Further, to ensure this wasn't some aspect of my camera or artificial lights, I repeated the process on a photo from last year taken with a different Fujifilm (X-T2 vs X-H2s) camera. The results are identical—bluish-reddish vertical banding in the deep shadows (not checkerboard like in your results).

  

The new and old files are here:

https://1drv.ms/f/s!AjAOXAIjC1XCgaMA235qbZFycHrnZA?e=ct6D9y

[DdeGannes]

Screen capture from Win 11  system LrC 13.3.1

[Ian Lyons]

@PvlKnn 

I've moved the thread to the 'Bugs' forum

@Rick Spaulding -  Please ask QE to investigate issue discussed in this thread

[DdeGannes]

Screen capture of denoised image and raw image with edits, viewed in the Library module compare view.

[johnrellis]

I get different results than Ian on LR 13.3.1 / Mac OS 14.5 -- not a grid of squares as he got but vertical purplish and reddish banding.  

But on LR 13.2 / Mac OS 14.5 and LR 13.3 / Windows 11 (no GPU in a virtual machine), I got much different results, with no strong banding, similar to Ian's result on 13.2.

I think the distinct banding / gridding and the variation of results between LR 13.3 and 13.2 indicate a clear bug.

- LR 13.3.1 / Mac OS 14.5:

- LR 13.2 / Mac OS 14.5:

- LR 13.3 / Windows 11, no GPU:

- LR 13.3 / Mac OS 13.3.1 [sic], no GPU:

[DdeGannes]

The screen capture is a comparison of the raw image with Denoise done on my Win 11 laptop on the current (up to date) build the GPU has limited acceleration, and the origin Denoise image supplied by the Author.

 My laptop does not appear to have the experiance of the Author's Mac.

[DdeGannes]

I had a look at the Camera Metadata to see the Camera settings for the RAW image 1/40 sec, f5.6, ISO 680, Exposure bias -2.67. 

What effect if any will the Exposure bias have on the noise in the image?

[PvlKnn]

The exposure bias (i.e., the exposure compensation) applied to the photo would have no additional impact on noise versus what we'd see had I used the same aperture, shutter speed, and ISO values in manual exposure mode. It's just a bias versus the camera's standard exposure index for that frame. 

[DdeGannes]

The reason I asked the question of "Exposure bias" is because of the aparent significant noise in the underxposed areas of the image. Looks as if the image was captured in the equivalent of  ISO 10,200 due to the 2.67 bias, just my impression. The image on the left is from the Library module compare of  Denoised image vs the original raw image at 100% with edits applied.

The image taken in Library module compare mode 100%. L to R, Denoise image vs raw image before Denoise

[PvlKnn]

This bug goes beyond this particular image. 

I just want to clarify: the photo was captured at ISO 640. The stated exposure compensation was applied to preserve highlight detail. There is no ISO 10,000 equivalency. The image appears as intended when you open it in Lightroom using the default settings. There was no intention on my part to process the photo with a +5 Exposure adjustment in Lightroom. I discovered the bug by accident. The RAF was denoised, and I was playing around with the resulting DNG when I noticed some odd colour artifacts, so I looked deeper into the shadows...

[DdeGannes]

@PvlKnn , understood, I was just making an observation no offense intended.

[Jao vdL]

Clearly a bug indeed. Can reproduce it with the files linked above. Appears that it happens only with files from this camera which is not uncommon. I can't make it do anything like this with any of my extreme ISO images.

[PvlKnn]

Again, I must insist this isn't an extreme ISO image (640). It's a photo with extreme exposure and shadow adjustments made in Lightroom. 

I was able to recreate it using a different and much older fujifilm model (X-T2). That raw file is attached above in one of my replies. 

Perhaps it's an artifact specific to X-Trans sensors? Frankly, I wouldn't be surprised by that given the decade of complaints with how CameraRaw treats those files. 

[Jao vdL]

P.S. interestingly Topaz also completely messes up this file. Using denoise and sharpen and dng output in Topaz PhotoAI gives this:

I've noticed these blocky artefacts in Topaz' output in other low light images before so no surprise but we are really at the limits of the current technology here apparently...

[PvlKnn]

I'd argue your Topaz example is the preferable version. I'm not expecting miracles in such deep shadows—it's mostly noise. It'd be nice not to have dreamt-up colours, though. 

[Jao vdL]

I don't know. The blocky stuff is also present in lighter areas like above the display case where there is clearly resolvable material. They're both pretty bad though and not really usable. The color blocks are extremely objectionable indeed

[joel8x]

I've been experiencing this same issue. I see that I can still downgrade to 13.2 - Has anyone tried this yet, and if so have you had any issues with the catalog?