Actually, maybe I should try LMMSE when enabling noise reduction…
In this case using IGV instead of LMMSE gave me a better result. But this can only be seen when zooming in to at least 200%. Being able to use and tune wavelet denoise made the biggest splash. Not sure if an (semi) automated wavelets denoise option would be feasible for or in the spirit of Filmulator.
In the end both are great for high ISO images, though.
I’m actually thinking now that the reason the colors differ so much might just be not-quite-correct black level subtraction.
Looks like RawTherapee has a black level offset for the G9, and I’m not using that value from camconst.json.
Sorry it was DT…xmp meta can be had by loading the jpg as a sidecar…
1 Like
NIGELSWAN GirlsHockey 1233001.dng.xmp (14.7 KB)
NIGELSWAN GirlsHockey 1233001.zip (7.6 MB)
(I can’t upload .jxl files, so I zipped it).
Full-res jpeg with mozjpeg-q85 to have something to preview.
This is basically simple, no-nonsense darktable:
- exposure up quite a bit
- color-calibration, trying to find a nice looking white balance is a bit tricky. I got a nice hit somewhere around the mouth guard. The average of the whole image seems better than the as-shot-in-camera to me, but this is all subjective.
- filmic, there isn’t really any real high dynamic range, but I basically use it for everything these days.
- local contrast with default settings
- local contrast but with contrast 4%-and-high-detail-overdone bilateral, but with a soft luma mask and final opacity around 50%. This gets some smaller details out in the field surface.
- I think I enabled ‘sharpen’ with default settings in the end.
But I did my usual ‘cheating’ thing → I started with a DNG saved by Dxo Photolab, with only optical corrections (vignette, distortion) and noise-reduction + sharpening, all on their auto settings (but their latest and greatest noise-reduction).
As a rule of thumb, if you get close to the detail/noise level that this does, then you can call your result ‘close to market-leading’ 
.
There is a bit of depth of field in this picture (m4/3 or not) so judging ‘details vs noise-reduction’ or ‘this might be slightly out of focus’ can be tricky by looking at it.
3 Likes
This is in the “Panasonic DC-G9” section in RT’s camconst.json file:
"black": 15, // 15 is BL offset. dcraw/RT reads the base black from Exif and calculates total BL = BLbase+BLoffset
Yes, I saw that.
Currently I’m working on making sure I don’t mix up actual black levels (like 512 for Sony) with offsets like the Panasonics, because it’s stored as just “black” for both.
My DxO DNG into Filmulator 0.11.2rc1.
I tried to match exposure (0.8333) and raising the shadow-brightness a bit (315) to match the Darktable try I did earlier. I lowered ‘film area’ a bit from the default (428), to get some more pop into the image. Oh, and I set white balance on the same spot (yay for white-balance picker!)
I like the red of the shirt more in the filmulator version.
3 Likes
DxO’s noise reduction is incredible.
Before anyone wracks their brain too hard about the magenta shadows from what I understand it is a known issue that the G9 has a magenta shadow tint. At least that is the info I found when doing research before buying mine. It appears both at high iso and if pulling up shadows. Flaw in the camera from what I understand. I find it not noticeable in normal situations so fine but if pushing limits like this photo you will probably need to color correct for it. Maybe RT is correcting for the issue Panasonic never fixed in firmware.
1 Like
Here’s a new version with proper(?) black subtraction (only 95% of the correction RawTherapee prescribed, because I think it overcorrected it), LMMSE, and adjusted noise reduction parameters.
It sure isn’t anywhere near the caliber of DxO Prime, but I think it looks quite good.
Interestingly, I’m finding that nlmeans doesn’t know what to do with the output of LMMSE; it somewhat ignores the finely textured luminance noise and mostly hits chroma noise. I still needed to use the directional pyramid denoise from RawTherapee to remove some of the chroma blotchiness, though.
Non-NR parameters:
- Tint: 0.9224 (I thought it was off with the camera WB)
- Shadow Rolloff Point is back to 0 because I didn’t need to hide the black subtraction issue.
- Drama 50.7 still
- White Clipping Point 0.631
2 Likes
This is a good test image as there is a good amount of both luma and chroma noise…
1 Like
@blj yeah I have noticed that, figured it was a quirk of the camera. It also is not consistent when doing multiple exposures (like 10s exposures when light-painting) even when everything is set to manual fixed settings. Most will be magenta tinged, some will be green heh.
@jorismak that is unreal crispness from what that started as! I’m guessing Dxo Photolab doesnt run on Linux, from looking at their site.
Of course, if I happened to be doing this for a living I’d get something like an A7R4 with BSI sensor for a bit of a leg up. But I am super impressed with this camera, bought the G9 second hand with 4K shutter count, so to me it’s like new!
I am wondering if I should have exposed to the right and pushed ISO to the max 25600 
I’ll see at another game.
In light limited situations there’s little benefit to ETTR. You’re well above the point where raising ISO helps signal-to-noise ratio.
1 Like
that was my suspicion yeah, it looked a lot noisier in the camera preview for no real gain (shutter was high enough), so moved it back to 20000.
Wow, I took a look at the pp3 and there are a hundred fifty three parameters to the wavelets tool, and that’s not counting multiple curve points per parameter.
It would certainly need to be boiled down to one or two parameters, I think, like I managed with the chroma NR (which is just one strength parameter in Filmulator).
I think I’m happy with nonlocal means though.
Yeah, the complete wavelet module in RawTherapee has a ton of (fine-) tunable options.
Assuming that the base settings can be easily automated (based on small/average/large MP RAWs) you’re still left with 15 or so denoise/refine related options. Some can be left at their default settings but that still leaves 5 (6?) that need to be tuned.
And to be honest: I did also use wavelets’ Edge sharpness to get some of the details back. And that one is tricky to set up!
I think that might fit Filmulator better then a wavelets based denoise option.
1 Like
darktable i warmed the image even though tehnically it would be quite cool and green. i left more noise in to keep some of the texture and what detail i could
NIGELSWAN GirlsHockey 1233001.rw2.xmp (34.2 KB)
1 Like