Output from Denoise AI significantly different than input

I’m not sure if this is a bug, if I’m doing something wrong, or if this is expected behavior from Denoise AI… I don’t RECALL ever seeing results this drastically changed from previous versions, but if it’s normal to have such a drastic change it would be good to know.

I couldn’t find any discussions of the output from Denoise being drastically different than the input, but I’m seeing that, and I was notified by another Denoise AI user this morning that he too is seeing this from his Nikon RAW images.

I took a standard, moderately High-ISO (8000) RAW exposure from a Nikon D500 and made only basic exposure corrections in Lightroom – exposure, black and white points, shadows and highlights adjusted. White balance as shot, profile normal camera standard from the input, no Presence, HSL, tone curve or sharpening adjustments. Luminance noise reduction set to zero. No significant adjustments.

I generated a .dng file using the Adobe .dng converter, so the .dng file is virtually identical to the RAW image. I also exported a .dng file from the original RAW using Lightroom’s Export to see if that would make any difference. All three source images are essentially identical.

I started Denoise AI standalone, and added the RAW and .dng images.

I processed them through the Denoise AI method at the Auto values: Settings were N19, S37 with recover detail at 5. (just in case there’s a question of using Low Light instead, I did. I ran a second time in Low Light at the Auto settings, and the outputs were the same)

I processed the images with the output as .dng.

I then processed the images with the output as .tif.

Both the .dng and .tif outputs differ drastically from either the .dng or .nef input images.

Both the .dng and .tif image are identical for a given input. So for the output from the original RAW image, the .dng and .tif are the same, but drastically different than the input image.

The same holds true for the .dng and .tif images derived from the exported or converted .dng.

I examined all the images in Photoshop, using histogram statistics (mean, Std Dev and median) and threshold values (black point, midpoint and white point).

(unfortunately, I have no idea how to display a table in here)

File Black pt Midpt White pt Mean Std Dev Median
Original RAW 1 128 248 103.32 49.29 100
DNG from RAW 9 121 230 60.48 27.25 56
TIF from RAW 9 121 230 60.48 27.25 56
Exported DNG 1 128 248 103.32 49.29 100
DNG from DNG 14 117 241 99.11 35.53 95
TIF from DNG 14 117 241 99.11 35.53 95
Converted DNG 1 128 248 103.57 46.77 101
DNG from DNG 14 117 241 99.11 35.53 95
TIF from DNG 14 117 241 99.11 35.53 95

It appears that the output from Denoise AI has significantly darkened the images and truncated the tonal range. The worst results appear to be from either .dng or .tif processed from the original RAW input. But all the outputs from Denoise AI seem badly compromised.

In theory, I’ve uploaded three images as small jpegs - the original raw, the .dng and .tif outputs from Denoise AI. Unfortunately, it appears I’m not allowed to upload images to this forum.

Has anyone seen behavior like this?

Try using TIF instead of DNG to see if that makes a difference.

I have experienced similar behavior with Denoise AI and DNG files. To get a handle on this I created TIFF and DNG files from my Nikon NEF file using Lightroom. The histograms were essentially identical to that of the NEF file. No processing of the NEF had been done except lens correction upon import. I then exported the converted file from Lightroom to Denoise AI specifying “Original” as the file type. The histogram of the returned TIFF file looked essentially the same as the TIFF that was sent to Denoise AI and the original NEF file. This was NOT the case with the DNG file. The histogram was shifted way to the left, i.e. the image was darker.

“Try .tif” isn’t an answer. DNG is a valid, listed output. If you read the original post, I state that I ran both .dng and .tif files. And on a normal RAW image with only the most basic exposure adjustments, both the .tif files and the DNG files suffered similar problems with exposure and loss of tonal range.

If the tool is going to have DNG as an output format, it seems like it should create output that doesn’t deviate significantly from the input, and that is recognized as a DNG by a program that opens it?

It is NOT a DNG in the sense of converting a RAW file and, just so you will understand the theory behind DNG files from Topaz AI products they are Linear DNG.

The Linear DNG format is scene-referred which means that it will show the colors directly in the image, rather than how your display chooses to show them. A scene-referred image is an image where the image data is an encoding of the colors of a scene (relative to each other), as opposed to a picture of a scene. In a picture, the colors are typically altered to make them more pleasing to viewers when viewed using some target medium. Linear DNG will look washed out , due to this! Please, understand that Linear DNG conversion is meant to provide more control over your image editing, rather than make the image more pleasing.

The “linear” part refers to the fact that these R,G and B values have no tonal manipulation applied and hold RGB image data, but it isn’t constrained, and 4 or more colors are allowed.

And not all programs can process Linear DNG.

OK. Sounds like generating DNG files from Topaz may not be useful with Lightroom. I’m not sure why the .tif files look so bad in this particular case - they match the DNG files. But, generally when called from Lightroom the tif output looks OK, so it isn’t a big problem.

I received an answer from Topaz support…

This is a known conflict with some raw files that aren’t technically
supported by our raw processing library. Occassionally you will see
magenta behavior as well. For any errant raw conflicts like this, it’s
almost always a result of the camera model not being supported yet. We
are adding over 1000 models in coming updates to our software, but
that is still being worked on. For now, for any image that has this
type of behavior, your best solution is to convert it to a TIFF prior
to importing, this should still give you a data rich raw-like photo to
edit, but it will play nicer from software to software (not just ours)
to help avoid conflicts like this.

Look for a permanent solution to this for most camera models in the
coming months.

So, it’s a known problem. It appears the 4-year-old Nikon D500 is not yet “technically
supported by the Topaz raw processing library”. As apparently is not the 6-year-old Nikon D810. I don’t even want to TRY this with the only 3 1/2-year-old Nikon D850…
It seems odd that the D500, which is sold with an emphasis on sports and wildlife photography, both styles that rely heavily on high-ISO imaging, wouldn’t be supported, but maybe at some point in the future.

Really? Nothin’?

So what is the issue, you posted the answer you received from support 2 days ago saying your camera isn’t supported but they are working on a solution but in the meantime they have provided you with a workaround.

The library used is LibRAW 0.17 … and is in the process of being upgraded.

A little surprised that no one reading this raised any question… Maybe I’m too used to Adobe.