Possibly, or noise may have been added ex post using a sensor noise profile.
The point is that you get back the kind of textures from the training set. May not be what one is after.
1 Like
I’m game for that. What would you consider to be high ISO, given that at some point extreme ISOs become futile under any condition? FYI, I shoot with a Canon R7 crop sensor body.
For this test I would go with 25600, that still retains a lot of information but is beyond the point I would shoot at with current APS-C sensors.
Please shoot still subjects, indoors (no wind) with a tripod so that the “slow” version is sharp too. Thanks for doing this!
EDIT textured fabrics or plush toys make ideal subjects for this test.
1 Like
… and remote release (and setting change) as well 
2 Likes
Some Sigma SD10 images to play with (X3F raws available on request).
Shots at constant exposure up to 25 sec, f/22 in 95F ambient.
These files are licensed Creative Commons, By-Attribution, Share-Alike.
Here’s a set shot with the Canon 5D mk. II:
2025-03-05t23-04-23_1990.cr2 (25.5 MB)
2025-03-05t23-05-46_1993.cr2 (29.5 MB)
These files are licensed Creative Commons, By-Attribution, Share-Alike.
F11, ISO 100, 10" and ISO 6400, 1/6" - what the camera considered correct exposure.
I also shot at ISO 1600 and ISO 3200, if they have any interest.
Basic darktable edits (capture sharpening, lens correction, exposure, saturation, no denoising) for reference:
I don’t have any of the fancy denoisers, so I’ll leave that to someone else.
2 Likes
All, please crop your images to a region of interest and apply at least a little bit of jpeg compression.
Sorry about that. I can delete my post and redo it in the morning, if you wish.
** I went ahead and took it down **
Hello @Donatzsky
At home, I run both RawTherapee and DxO Photolab 8.3.1 on Windows 11
DISCLAIMER:
DXO is NOT freeware, you must pay in order to run it perpetually (unless you choose to try its demo version for 30 days…). In addition, it does not run on Linux.
I opened your image with DXO and denoised it. Without any personal settings of mine changed, only the one set by DXO by default for this RAW. To run this test I have applied its latest denosing algorithm.
Here is its GUI:
Here is the jpeg (80% quality) exported (ISO 6400), denoised by DXO (default settings):
1 Like
Unless @Dave22152 has access to all non-free denoising engines that could be interesting to compare and is willing to generate all images, it would make sense to post the two RAW files.
Just a general comment that crowning the best will be impossible because
- Skill levels in different software differ
- There’s an element of aesthetics and subjective preference involved.
That is very good indeed. Less noisy than ISO 100, without looking artificial. Some of the text has a bit of a wobbly look, but that seems to be due to the noise effectively changing the shape. Compared to ISO 100 you also lose some fine texture, but that was already obliterated by the noise, so that’s to be expected.
2 Likes
I think the objection was my posting full sized results at minimal compression and not the raw files I posted as well. I get that and I’ll try to repost with smaller cropped regions to compare them the problem is that significant compression can make it difficult to compare noise reduction results
I have several of these programs, so I’ll try to repost with Topaz, DxO in addition to Rawtherapee and DT.
@nosle , from my perspective it isn’t about crowning which is the best, but to give users a better idea of how each program works as well as their drawbacks to let them decide which software is best for their own needs.
2 Likes
Sharing the raw files is up to you. I just want smaller jpegs. For purposes of comparing noise reduction, it seems like a good idea to not resize, but rather crop to a region of interest.
My comment wasn’t for @Dave22152 specifically, it was for everyone
1 Like
I totally get it, and I didn’t feel like I was being singled out! I created a new thread with more reasonable file sizes
2 Likes
I agree there really is a point where “AI” based models become problematic in defining what is signal and what is hallucinations!
As for if we are at the limit of denoising performance, I want to argue for the case that there is plenty more signal to noise ratio to be extracted from images using “classical” methods that do not hallucinate. Here are a few things that isn’t done today in the open source tool set:
- Average blacks instead of clipping! Clipping “negative” pixel values will skew denoising results and likely result in loosing saturation in noisy pictures. dt unfortunately clips pixels in the demosiacing step.
- Work on edge aware denosing that lets more noise trough at edges, similar to what a lot of people already do but relate it to the expected noise variance.
- Local color components instead of RGB or luma/chroma. There are huge compression possibilities by basically performing a PCA locally. Higher compression means that we can denoise better.
- Static dark frame subtraction! This is the same as having a per-pixel calibrated black levels for your camera. Potential for slightly better colors. *Static as in profile once in “normal” conditions, not for every shot as the astro folks do.
- Combine demosiacing and denoise! Denoising works best if we can assume that the measured signals are uncorrelated, demosiacing destroys this assumption by interpolating data using correlation between the channels before denoising
I’m trying to learn more about the subject and I’m feeling optimistic about that we haven’t reached peak performance for classical methods yet. 
10 Likes
I deleted my post as I posted to the wrong thread.
I looked into combined demosaicing & denoising and listed some papers I found interesting at
but treating this as an inference problem is probably too costly.
Working on (local) principal components could be fruitful, but it may be tricky to 1. achieve consistency with neighboring pixels (which will have different principals), 2. combine with demosaicing.
If you experiment with these, please keep us posted on this forum.
2 Likes
Nice that you also have observed the joint demosaic and denoising problem! I haven’t come that far yet, mostly spending my small experiment hours on compression using both local pca and spatial methods. Still mostly failing forward so not very post worthy yet 
2 Likes