That’s looking pretty good for the type of noise reduction it is. You’d have to go to the frequency domain to pull out the texture of the pencils I think.
Thanks for the hint. If you have any tips or examples, that would be great. A little help goes a long way for me. Be gentle though. I have trouble following your *.gmic file. 
For comparison, here is the result from denoise_patchpca using default parameter values. Takes 100x longer.
I think a more reliable comparison is having more pictures to go on by. denoise_patchpca doesn’t seem to work for that image. It reminds me a bit of AI art error where you see those weird patches.
You know, I think I may provide you a thorn fractal noisy image with variety of colors tomorrow or two. I feel like doing that, but it’s late here.
Even I have trouble following my gmic file 
Here is the result from my Noise reduction filter, which is much more complicated.
The main difference is your result has some colour smearing which is always a problem, especially if you are using a luma/chroma colour space like Lab or Ycbcr.
With regard to tips. Although I haven’t even looked at your gmic file, so I might be telling you something you already know.
There are two avenues you could pursue:
- Consider the guide image you are using. You can manipulate the guide image to improve the noise reduction. For example, my filter has the ‘soften guide image’, which is a small radius blur on the guide to knock back impulse noise. You can do all kinds of manipulation to the guide and see if it helps. For example, you could try improving colour edges by increasing saturation at edges (or decreasing saturation away from edges) in the guide.
- Use your filter in multiscale. Your result suffers from low frequency blotches, which are easy dealt with on a scaled-down version of the image. You may find you need to do less noise reduction on the full-size image if you are dealing with blotches on other scales.
1 Like
Here is a test image I use for noise reduction tests. It from a humble 6-megapixel D40, which is about the right amount of pixels when you are working with slow algorithms.
1 Like
Sometimes it is best to leave it alone for a while so your brain can come up with a solution on its own. If you keep looking at the problem all you see is the problem and not the solution.
Hi!
I am afraid I don’t have a lot of picture that are at the same time noisy, with details and structure, and with wide dynamic range.
You can find some of my noisy images here, maybe some images will be ok for you (I think DSCF7337.RAF has all the criteria, but that might be the only one): https://drive.google.com/open?id=1EaVApc2U-PlvwCeZT3A-xkqePOeA282f
Addition processing yields sharper structures: noisy, previous, sharper
Sharper looks sharp but is more abstract. With one greater unit radii.
:
@s7habo Is that a cheerio?
Also reminds me of rough cuts found on a palisade, marshmallow stick or stone age tool.
Status update
– Fixed a couple of afre_gui1* copypasta bugs.
– Explored a bit of multi-scaling with @Iain. Wasn’t impressed by what I saw. Rather than remove the low freq splotches, introduced more problems (patterns, haze and halos). Back to reading papers and pretending to understand them.
1 Like
My English is so bad that I don’t know that animal. That’s supposed to be beaver. 
Was referring to the American breakfast cereal (bottom right).
On the left, there is a mini stargate.
1 Like
Seem like you’re getting to the point where you’re solving it, I take it?
Not by a long shot. 
In fact, the method I used on sample pencils noise 40 cut 0,255 destroys @Iain’s sample image.
This much simpler method is more consistent.
Original
“Denoised”
Noisy
“Denoised”
1 Like
I think that’s about the best that is possible. If I attempt to denoise using my own brain, the denoised result is how I envision it. However, on the right-bottom corner, it’s not how I envision it though.
Update
1 afre_compare Added a tile mode. axis is now called layout to accommodate it. Several more to do items for this command.
afre_compare:
crop={ 1=once | 2=twice },_layout={ x | y | z | tile },_normalize={ 0 | 1 }
Crop then compare selected images interactively.
Default values: 'crop=1', 'layout=x' and 'normalize=1'.
...
As usual, wait several hours before updating your commands.
2 I am going to let go of my “denoise” efforts for now. I use scare quotes because the motivation was to see how far I could get with my afre_gui* in a minimalist way. It is mostly because my tank has been empty for some time. More time + effort ≠ success.
1 Like
Awesome update
1 New afre_sdpatch Similar to variance_patch but an improvement. Was about to call it “afre_lsd” for local standard deviation but you know.
2 Feature afre_edge now has a parameter method that allows user to choose between afre_gnorm and afre_sdpatch.
As usual, wait several hours before updating your commands.
2 Likes
Regrettably, when I refreshed my system, I forgot to back up user.gmic. About a year (and a half) of R&D and private commands gone. 
Oh well, no use complaining about it. I can rebuild based on my other *.gmic files.