Show me how you denoise

FAb7D · May 8, 2023, 8:28pm

Thanks, it works, I had not realized that there were 2 versions trougnouf and daniezest… I thought daniezest was an English word I didn’t know
So I’ve used trougnouf with the model1 of daniezest:

Thank you!

sillyxone · May 9, 2023, 12:42am

Thanks, it works, I had not realized that there were 2 versions trougnouf and daniezest… I thought daniezest was an English word I didn’t know
So I’ve used trougnouf with the model1 of daniezest

Glad that it works. Depending how much memory your GPU has, you might run into out-of-memory error occasionally. For my 6GB RTX 3060, I added an export statement before calling nind-denoise:

export 'PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512'; python3 denoise_image.py  ...

Your image is scaled down so I can’t tell for sure, but looks like it’s still quite noisy. The model was trained with almost-raw data, right after demosaic with only brightness adjustment. Thus, you need to apply nind-denoise to the full-size image with no denoise, sharpening, scaling, rotating, … or any operation that can alter the original noise pattern.

That’s why I have a Python script taking advantage of darktable-cli. It copies the darktable history-stack/sidecar into two copies: applies only demosaic/brightness/sigmoid/filmic in the first copy to export into a TIFF, runs nind-denoise on that TIFF, then imports the output from nind-denoise and applies the rest of the history stack. Otherwise, with the noise pattern altered, the output from nind-denoise likely ends up with weird artifacts.

github.com

hqhoang/nind-denoise/blob/darktable-cli/src/darktable/dt_nind_denoise.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
@author: Huy Hoang

Split the darktable export into two stages to inject nind-denoise into the
history stack

"""

import os, sys, subprocess, argparse
from argparse import ArgumentParser
import configparser
from bs4 import BeautifulSoup
import copy

notouch_ops = [
]

# list of operations to be moved to second stage

This file has been truncated. show original

ThomasO · May 9, 2023, 1:09pm

I am currently also hunting the king fisher and and stumbled across this photo.
I am fighting with the same problems…noise, details,…
My result with DXO PureRaw3 - ART - Topaz Sharpen AI.
I do not want to step on anyone’s feet but the for me the shot is really good and enough potential for tweaking…with available tools on the market.
Critisism welcome

piratenpanda · May 9, 2023, 3:39pm

Thanks, nice to see how proprietary tools fare. No option for me but interesting nevertheless. I find it a tad too sharp to be honest.

FAb7D · May 9, 2023, 8:19pm

Yes, so I just tried demosaic and brightness from Rawtherapee. I saw (trying) that it does not read 16 and 32 bit float, so I exported a 16 bit, rgb, tif. I tried demosaic with IGV (good for high ISO) and AMAZE+VNG4. It seems to me that using nind-denoise is a bit better AMAZE+VNG4 at least as a starting point for subsequent processes.
Anyway it looks very good to me. Convenient to have it on Darktable … I’ll have to learn how to use Darktable

sillyxone · May 9, 2023, 8:53pm

Yes, so I just tried demosaic and brightness from Rawtherapee. I saw (trying) that it does not read 16 and 32 bit float, so I exported a 16 bit, rgb, tif. I tried demosaic with IGV (good for high ISO) and AMAZE+VNG4. It seems to me that using nind-denoise is better AMAZE+VNG4 at least as a starting point for subsequent processes.
Anyway it looks very good to me. Convenient to have it on Darktable … I’ll have to learn how to use Darktable

Watch out for artifact when you use combo of demosaics such as AMAZE+VNG4, since the noise patterns are no longer uniform (AFAIK, AMaZE and VNG4 are applied at different detail levels). I ran into artifact when using Markesteijn+VNG, and had to switch back to just Markesteijn (for X-Trans), 1-pass or 3-pass doesn’t make much difference.

I think the existing dataset was demosaiced with AMaZE (for Bayer) and Markesteijn (for X-Trans), so you’d get the best result sticking to those two.

Both darktable or RT would work fine with nind-denoise. It just happens that I’ve been using darktable due to GPU support and familiarity, and darktable-cli is scriptable.

FAb7D · May 9, 2023, 9:01pm

Ok, thank you very much!

qmpel · May 10, 2023, 9:56am

First of all, @piratenpanda, I love this picture. Thank you for sharing.

I had great fun playing with the different denoising algos - thanks to @Terry for bringing some lesser known modules back on the table.

Here is my result (only DT)

and a bit cropped

It took a while to get to the result and for denoising I would love a more effortless solution in darktable - but on the other hand, I often dislike the results of the one click AI based algos (beeing to smooth and like plastic).

Let’s see what the future brings…

edit: Not sure if I uploaded the correct file.
Also, I pushed the denoising more, than I normally would - actually I like it a bit grainy…

Underexposed · May 10, 2023, 11:41am

I especially like what you did on the cropped version. Would you mind sharing your sidecars?

qmpel · May 10, 2023, 11:51am

Sure, let me find the sidecar.

I didn’t include it, because it is kinda messy. Lots of masks an modules. But in the end, I’m happy with the result. Could always be better and the masks could be more precise… but it is what it is

This is my current version, its a bit brighter - probably still not enough. I feel it looks a little diffrent, viewed from the browser - a bit darker.

151A2902.CR3.xmp (50.5 KB)

killed chroma noise with denoise
reduced luma noise with denoise nlm
crushed the background noise with contrast eq
denoised the bird with astro denoise (lol), could probably have used denoise here also
a steep tone curve for the birdy
added color with color zones (for the bird)
added a bit grain back, so it doesn’t look to artificial

Thanks for asking.

qmpel · May 11, 2023, 7:24am

So … sorry for highjacking this thread, but I don’t think it would be worth a new thread. I just wanted to thank you for mentioning the astrophoto denoise, because meanwhile I integrated that module into my workflow for DIA positive restauration and I am quite happy with the results.

Here a little example how it works in this case (the source image is a macro photo of a DIA positive, is it called like that?!).

In the upper left you can see the typical grainy noise in the original. Upper right is trying to denoise with Denoise NLM, to get an effect in the face its really crushing the background. In the lower left is the astro denoise with moderate settings - i think it handles the noise really well, before I was always struggeling in this situation.
Lower right adds a tiny bit of contrast with the contrast eq and also a subtle bit of fine grain.

Terry · May 11, 2023, 7:33am

what is denoise NLM? I might know it by a different name or have overlooked it.

BTW, daktable has many denoise options and the best ones can vary between images. So be a little flexible in which modules you use.

qmpel · May 11, 2023, 7:36am

Hey Terry, I meant “non local means”. Of course the noise varies between images, but here I’m talking about the specific grain I get in digitized DIA positives and this type of noise seems to fit well with the astrophoto module.