New Capture One AI denoising

[myotis]

Just released in public beta (No NDA) is C1's AI noise reduction. Initial tests look very good and below is one of my initial tests at 12800 iso.

On the left is DXO Prime 3 (latest version PL9.7 or PR6), in the middle is C1 default non-ai (default) noise reduction and on the right is C1s AI noise reduction. All at 100% view


As noise reduction results are not only specific to software but also to individual files, it's difficult to draw any conclusions. I've also done a bit of editing on the two noise-reduced files, trying to match their appearance. Which is the other issue with comparisons because the more time you spend tweaking the results, the smaller the differences between them.

I reckon DXO is still looks as if it's pulling out more detail, but it also has a "flatter" less "3d" look than the C1 result. It's the same conclusion I came to when comparing Adobe AI Denoise and DXO, and in so far I think I prefer the C1 results.

I haven't tried any really high ISOs yet, and it was with higher ISOs that, overall, the detail enhancements provided by DXO outweighed the more natural look of Adobe AI Denoise. However, at the moment, the C1 implementation is looking pretty good.

It works on the base RAW layer, and disables the original denoising tools. But you can still add a layer and use the old tools to further adjust luminance noise removal, and detail enhancement, selectively.

You only get one slider, which adjusts the strength of the noise reduction, and allows you to balance enhancing detail vs reducing noise.

I haven't timed it, but I reckon the AI noise reduction takes about the same time as DXO. The AI noise reduction information is stored in an XMP file, which increases file storage by between half and the same amount as the original file, Which as I remember, is about the same as the increase in storage needed by Adobe DenoiseAI since they stopped using a DNG interim file.

The new DXO compressed DNG format, which shows no loss in quality, produces DNGs that can be as much as half the size of the original RAW.

It's been a long wait, but at the moment it's looking good for C1.

 

[Pound Coin]

Oooh, that is interesting. No more round trips to Topaz would be good.

I wonder if they are working on sharpening too? (Unless I've missed an improvement)

 

[myotis]

Oooh, that is interesting. No more round trips to Topaz would be good.

I wonder if they are working on sharpening too? (Unless I've missed an improvement)

Well, one of the interesting things about the denoising is that you can now push the sharpening much further, and the AI denoising itself is adding a bit of capture sharpness.

Having said that, what do you find lacking with the sharpening?

 

[Pound Coin]

Well, I find it lacking in finesses and not easy to control. It may be that I'm just used to the way that Topaz Photo deals with it.

It's quite possible that the noise reduction in C1 will work well with Sharpening, and I'll have a play when it is released!

 

[Barney]

You can really see the difference in the reflection in the eye.

Is that with or without sharpening?

 

[myotis]

You can really see the difference in the reflection in the eye.

Is that with or without sharpening?

I did a bit of editing to get them both looking similar, which probably included a bit of sharpening or structure modification.

The noise reduced DNG from DXO also has "lens softening" sharpening added by DXO, but C1 treats the DXO DNG as a raw file and auto adds capture sharpening. So you need to remove all the sharpening in Capture One, and then bring slowly it back, until the sharpness on the DXO file matches the sharpening on the C1 denoised file.

Then you need to decide what the "correct" sharpening is, and iterate the sharpening on both files until they both look as if they have the "correct" and "matching" sharpness, but this means the actual sharpening settings will be different for each file.

Hopefully that makes sense.

There is a real problem with this type of comparison as it is useless to compare defaults, because that doesn't tell you anything about the full capability of the different programs, so to get something that gives a more useful comparison you need to edit them, which means you aren't really comparing like with like.

The differences in the eye could probably be fixed with a bit of local editing.

The key take aways for me are that the C1 offering is seriously good and worth looking at, and that there seems to be a fundamental difference in the way DXO builds the denoised image compared to the way C1 (and LR) builds one.

I couldn't "fix" the flatness and lack of 3d with the DXO image, and nor could I extract the extra detail that the DXO images seems to offer compared to the C1 image. This was the same experience I had when comparing DXO with LR.

 

[myotis]

Another couple of comparisons, this time at ISO 25600



DXO on left and C1 on right

As before I have adjusted both these pictures in C1to try to maximise results., and as always, there are multiple ways of processing them, including changing settings in DXO, I always use the default noise and lens sharpness settings in DXO, but reduced from 40 to 30 for noise and 100 to 90 for sharpness. Which minimises over sharpening and loss of detail from too much noise reduction. I then finalise noise and sharpness in C1.

With DXO I still have recurring issues with "areas" of softness, some artefacting, and haloes (but not as much as Topaz), and you can see that DXO example here as areas of softness and some haloeing around the whiskers: which I couldn't fix.

Although it isn't obvious from this crop. In the areas where DXO performs well, it seems to capture more detail and give a "smoother" result than C1, but overall, I am really liking the results rom C1.

From my ongoing testing, I've always concluded that although Adobe AIDedoise gave the most natural results, at higher ISOs and overall, DXO was still the best choice. I have found Topaz to be too unpredictable. However, these initial tests are looking very good for C1.