Denoise challenge - raw, jpeg, tiff provided

sluggy_warrior wrote:

Rightsaidfred wrote:

sluggy_warrior wrote:

FWIW, darktable with just filmic, exposure, color balance, color calibration, and tone equalizer. Export fed through nind-denoise, then GMIC's RL-deblur.

I see plenty of artifacts, but nind-denoise is still in its early stage, so I'm not complaining

Yes it is quite uneven. Parts are quite good, others are not.

The software needs to be trained. I've not understood, does every user need to train on their own, or is training centralized?

It only needs to be trained once, then you can use the trained model to apply denoise to new images.

The author trained his model with limited samples from X-T1 and perhaps a Canon DSLR. The plan is to add more samples, but there are other questions to be solved, too, e.g. which ISO to include, should dark/bright shots be included, ...

Also, the shooting technique to generate the paired truth is also complicated. We need to pair a clean and a noisy image together so that the model can learn how to transform noise into clean, the pair will need to be aligned perfectly pixel-to-pixel. Thus, shooting technique is critical, and doesn't work for moving subjects.

The noise pattern is also affected by the demosaic process (the denoise process is right after demosaic, before other sharpening/coloring). The current model was trained with Markesteijn, thus, doesn't work as good when switching to FDC, or Mark + VNG, for example.

And then, the denoise process is slow (8 seconds on my RTX 3060), that can't be integrated into darktable workflow. That's another problem to figure out with darktable team, too.

Many thanks for your explanations.

Very interesting.

I mean, it was a Master thesis - hats off.

Regards,

Martin