Are you satisfied with your denoise software?

[sluggy_warrior]

That last is a very good result. For what operating systems is it available? Scripting only or GUI?

It's written in Python, so it should run on most platforms. CLI only, no GUI (it doesn't have many parameters anyway). I've only tested it on Ubuntu.

GitHub - trougnouf/nind-denoise: Image denoising using the Natural Image Noise Dataset

Image denoising using the Natural Image Noise Dataset - trougnouf/nind-denoise

github.com

 

[CAcreeks]

That last is a very good result. For what operating systems is it available? Scripting only or GUI?

It's written in Python, so it should run on most platforms. CLI only, no GUI (it doesn't have many parameters anyway). I've only tested it on Ubuntu.

https://github.com/trougnouf/nind-denoise

Holy Monty, Python?!? I am amazed that a Python script is so high performing.

As you poste, "nind-denoise takes about 11 seconds on my 6GB RTX 3060 for a 24MP RAF."

DXO Deep PRIME takes almost that long on a typical PC, and XD almost twice as much.

Topaz "AI" takes about 100x longer to download their so-called noise models.

 

[Euell]

Yes, I'm happy with DxO PhotoLab v6.

Moderately happy with Topaz Denoise AI, but I would only use it on images that I can't process using DeepPRIME (i.e. non-raw images).

Noise reduction and sharpening for detail are two sides of the same coin. While Photolab 6 noise reduction of RAW files is excellent and I use it whereber possible, PL 6 lacks the best sharpening routines.

OTOH, DxO's lens profiles apply excellent capture sharpening. For creative and output sharpening, I rely on Lightroom.

For many files the sharpening of Topaz software and in some cases Luminar Neo will provide a better overall result. I print almost exclusively on13x19" and 17x22" papers and occasionally on 17x25".

Whatever capture sharpening is applied by DxO isn't a complete solution, as you note. In fact, DxO has a separate independent sharpening module, which strongly suggests that DxO understands the limitations of whatever sharpening is automatically applied.

Sharpening in Photoshop is not the same as Lightroom and Photoshop has no dedicated output sharpening.

Depending on the photo if I want the best quality sharpening, following whatever rudimentary sharpening is applied by DxO

There's nothing "rudimentary" about DxO's profile-based Lens Sharpness. It applies stronger sharpening in areas where a given lens model is known to be softer. Try that with any other app. In my view this is all that's needed for capture sharpening.

For creative sharpening, DxO's USM combined with U-point local adjustments is pretty versatile.

DxO lens corrections are very good, but print making usually requires additional sharpening and not just output sharpening. The photo below was trash coming out of DxO; the detail was smeared by the noise reduction.

The photo below was made with a lens that was incorrectly reported by the camera and thus DxO was unable to apply a profile for Lens Sharpness.

No,

Yes.

as with many lenses, DxO has no profile for that fisheye lens.

Nobody offers lens-specific sharpening for that lens. The example you're using to criticize DxO is an outlier. With most lenses, DxO does apply "additional sharpening and not just output sharpening". Adobe offers this feature for exactly zero lenses. In fact, nobody other than DxO is offering this feature for any lens.5

Oh, I wasn't criticizing DxO for not having a profile for THAT lens, but generally for many lenses, including, as I mentioned, almost all Voigtlander lenses, including the popular 50mm F2, which I have. Fortunately, that lens doesn't require significant correction. But, you are correct, DxO ignores manual focus lenses, even those that are reasonably popular.

But, noise reduction was applied in DxO, which is very good. Even if there had been a DxO lens profile, it would not have brought up sufficient detail for this shot, in my experience.

My experience is different. The combination of capture sharpening via Lens Sharpness and output sharpening via Lightroom's Export module is generally sufficient. If the lens is soft, as my FZ1000 MkII's can be, I'll apply fine sharpening USM in Lightroom. I could just as easily do that in DxO PhotoLab, but the masking slider in Lightroom's USM makes this tool a bit better.

Rest assured, following noise removal of high-noise files, the newer AI software sharpening routines can be FAR better that what is currently provided by Adobe, although I believe Adobe Lightroom may be looking into AI development for noise removal and sharpening.

It irritates me that despite repeated requests over the years, DxO ignores development of lens profiles for several relatively popular Voigtlander lenses (not this photo).

There are a lot of lenses hitting the market these days, including relatively obscure manual-focus ones. DxO has to devote its limited resources to the ones with the largest user bases. I've got a couple of Samyang AF primes for FE mount for which DxO does not yet have profiles. I've put in requests to DxO, but I'm not publicly trashing the company about it because, for the vast majority of my work, PhotoLab is stills hands-down the best tool for me.

Trashing? That's an absurd way to look at legitimate criticism, even if you personally disagree.

Your reference to DxO's sharpening as "rudimentary" struck me as trashing the company, not "legitimate criticism".

Not intended as a criticism at all. The intended reference was to basic or initial sharpening. Further sharpening, if necessary, is accomplished elsewhere in DxO or in another program.

Understood.

DxO has good noise reduction and some other features. The software isn't a one-stop solution

Nothing is.

and works best only for lenses in its database. That's an observation and opinion, not trashing or trolling.

I agree.

As I mentioned, I have had the DxO software for years. It has value, but is not an end-all be-all.

Nothing is.

Speaking of trolling, Jacques . . .

You're the one bringing it up, not me.

You misunderstood inferring a negative connotation to "rudimentary," which does have more than one meaning. Fair enough.

Overall, the comparison of PhotoLab with Lightroom is a mixed bag. DxO's Lens Sharpness function is quite sophisticated, not at all "rudimentary", and unmatched by any other RAW processor.

No doubt, the various levels of noise reduction available in Photolab are designed to play well with the initial sharpening applied in the DxO demosiacing process. Sharpening itself creates noise, of course, but there is logic in applying sharpening after noise reduction. DxO apparently recognizes this by including a sharpening module in Photolab, so that additional sharpening can be applied either before or after noise reduction. That sharpening module, however, lacks AI capability.

OTOH, its USM tool is fairly basic, lacking the masking slider found in Lightroom. OTOOH, one can apply masked USM via the U-point local adjustments. OTOOOH, Lightroom's output sharpening via the export function is quite sophisticated, and PhotoLab entirely lacks a function designed specifically for output sharpening.

While Photolab itself lacks dedicated output sharpening, DxO also produces the NIK plugins, which do include output sharpening.

Both apps have great strengths and relative weakness with regard to sharpening. I can't pick an overall winner. Which is why both have key roles in my workflow.

What detail there is was restored using Topaz sharpening. True, this was an extreme case

and an unrepresentative one

Not really. It was a noisy file.

, but I find additional sharpening using Topaz and sometimes Neo can make a big difference in prints.

That may be. But, in this case, you are not taking full advantage of DxO's capabilities.

And, how is that? Look, Jacques, DxO has its uses, which is why I have subscribed for years, but DxO has limitations that prevent it from being a one-stop panacea. For example, the algorithm of the shadow slider washes out contrast before you bring up shadows very far. Adobe RAW and Neo both work better for that. Bottom line is DxO works well as a single solution if you don’t intend significant manipulation and the file doesn’t need much sharpening. Heck, I can make due with Photoshop only for files I never intend to print.

, I will use Topaz or Neo. This is particularly effective for high ISO. I will also do an output sharpening for inkjet printing some photos with either NIK or the old Photokit output sharpener.

I used to use PhotoKit Sharpener for output sharpening until I read that much of the tech behind it had been shared by the devs with Adobe and incorporated into Lightroom's output sharpening in the Export module.

I use Photoshop, but not Lightroom.

Teatro Colon, Buenos Aires

Note: the lens was a fisheye, not 50mm as reported by the camera. While DxO provided good noise reduction, sharpening outside DxO (Topaz) was required to bring in detail.

 

[sluggy_warrior]

Since I did that test this morning, DxO has released a new version of PhotoLab, 6.4. This new release now allows the latest NR tool, DeepPRIME XD (eXtra Detail) to be used on Fuji RAF files. That recovers a lot more detail than the older DeepPRIME:

The name "extra detail" is too literal, I think I'm seeing details that wasn't there to begin with (e.g. texture on the pillow cover on the top-left). But then I realized I was too stupid to not have ground truth to compare against.

Anyway, started another thread with sample sets if you want to play. I include the clean ground truth shots this time :-(

https://www.dpreview.com/forums/thread/4704150

Hard to say. The tools tend to be very sensitive to GPU performance and drivers. My guess is the commercial tools are a bit faster than that.

Yeah, depending on the noise level, it's usually 8-11 seconds per image for me. At least unlike Stable Diffusion which requires at least 6GB of GPU memory, nind-denoise works fine with less than 4GB (the model isn't that big)

 

[newmikey]

TODO: C implementation that can be included in darktable, find lighter functional network architecture, reprocess raw data to train model at the beginning of pixel pipeline