Help for talk on AI Noise Reduction

[Chris R-UK]

I have offered to do a short talk for my camera club on AI noise reduction and need some help.

I have been using noise reduction software for some years now for nighttime soccer matches. I started with Topaz DeNoise AI but then got better results with the various versions of DXO Deep Prime. Recently I have played around with Lightroom's DeNoise but just find it too slow. For my application I start with extremely noisy images and, after batch processing (30-50 files at a time) in DXO, finish up with images that are acceptable for the web. They don't need to be high quality, they just need to show the action..

My talk is going to be aimed mostly at those club members who have never tried Topaz, DXO or any of the other AI noise reduction programs, but probably do have either Lightroom or ACR. So I am not going to do a detailed comparisons between the various options available, except for workflow, file sizes and processing times. But I do want to show them just what AI noise reduction can do and how it can change their photography.

My problem has been finding good examples of what AI Noise Reduction can do. I have 18 years of raw files in my Lightroom database but, until the last few years. I simply didn't take images at very high ISO settings - I either used flash (a lot), a tripod or I didn't take photographs. So I don't have any good examples of how AI Noise Reduction can revive old photographs. So, apart from my soccer shots which are quite extreme examples, I am either going to have to find some examples on line or I am going to have to shoot some very low light shots myself.

So the help that I require is:

• Examples of reviving old noisy images
• Examples of good quality images after AI noise reduction, as opposed to my lower quality soccer shots. Some bird photography examples would be good.
• Suggestions for the types of subjects that I can easily shoot myself that respond very well to AI noise reduction. Pets, flowers, hand held still life? I have lots of champagne bottles. :-D

I am also going to cover causes of noise, techniques to minimise noise in camera, e.g. ETTR, HDR, Image stacking etc., traditional noise reduction before AI techniques, and the benefits to your photography of not having to worry about your ISO settings. I will give some examples of how to use Denoise in Lightroom and I will cover software costs, file sizes, processing times, etc. Does anybody have suggestions of other areas I should cover?

 

[sybersitizen]

... apart from my soccer shots which are quite extreme examples, I am either going to have to find some examples on line or I am going to have to shoot some very low light shots myself.

So the help that I require is:

• Examples of reviving old noisy images
• Examples of good quality images after AI noise reduction, as opposed to my lower quality soccer shots. Some bird photography examples would be good.
• Suggestions for the types of subjects that I can easily shoot myself that respond very well to AI noise reduction. Pets, flowers, hand held still life? I have lots of champagne bottles. :-D

An easy option that comes to mind is to use some high ISO files from DPR's studio scene. You'll have nothing new to shoot, and plenty of detail for analysis.

 

[gscotten]

I probably would not dwell on processing times, as that is so hardware dependent. In my case, Lightroom AI Denoise is actually faster than Topaz Photo AI. In other extreme cases seen on this forum, members with different hardware report LR as an order of magnitude slower than Topaz.

My hardware is not high end: $1200 Dell G15 with i7-12700H and RTX3060L

 

[Zeee]

Adobe Denoise AI with ISO 20000.



Original





--
Don't Look Up.

 

[Chris R-UK]

... apart from my soccer shots which are quite extreme examples, I am either going to have to find some examples on line or I am going to have to shoot some very low light shots myself.

So the help that I require is:

• Examples of reviving old noisy images
• Examples of good quality images after AI noise reduction, as opposed to my lower quality soccer shots. Some bird photography examples would be good.
• Suggestions for the types of subjects that I can easily shoot myself that respond very well to AI noise reduction. Pets, flowers, hand held still life? I have lots of champagne bottles. :-D

An easy option that comes to mind is to use some high ISO files from DPR's studio scene. You'll have nothing new to shoot, and plenty of detail for analysis.

Great idea. Just done that.

 

[Zeee]

Here is a link.

https://photography-on-the.net/forum/showthread.php?t=1530101

 

[Chris R-UK]

Here is a link.

https://photography-on-the.net/forum/showthread.php?t=1530101

Thanks for both of those. The damselfly is an interesting recovery from a disaster.

 

[BertIverson]

Here is one I took a year ago RX10M4 - 1 inch sensor: (be sure to view original).
It convinced me to quit worrying about noise and sensor size !!!


Hand-held DxO PhotoLab 5 - Deep Prime NR -- 1/250 F 4.0 ISO 12800 FL 600

Bert

 

[mmartel]

So the help that I require is:

• Examples of reviving old noisy images
• Examples of good quality images after AI noise reduction, as opposed to my lower quality soccer shots. Some bird photography examples would be good.

Back in 2018 I was evaluating the Panasonic GX85 vs. Fujifilm XT20 and was shooting all kinds of random photos in different light with the respective kit lenses to compare the out of camera output and to see how they did when I processed them in my software of choice at the time (Lightroom Classic).

I grabbed this poorly-lit indoor shot of my son with the GX85 and humble kit Lumix 12-32mm f3.5-5.6. At 1/60s and f5.6 exposure, the shot is at ISO 10000. A portrait like this should, of course, be shot with bounced flash. But I didn't have a flash for either camera yet and it would take me a while before I learned the GX85 has a built-in flash you can tip back with your finger to bounce it off the ceiling. OK, on to the photos...

First, the out of camera JPEG. Your generally-expected chroma-noise-filled blotchy mess for a micro four thirds sensor with in-camera processing circa 2018.


GX85 SOOC JPEG (1/60s f5.6 auto-ISO @ 10000)

Second, as I processed it at the time with Lightroom Classic. Colors look a bit to warm to my eye now, but there is improved noise/detail here compared to the SOOC output. Much more realistic hair and much less color noise. Nonetheless, the hair still isn't great and there's still a lot of luma noise and it generally looks like a "too high ISO for micro four thirds" kind of image.


GX85 RAW processed in Lightroom Classic back in 2018 (1/60s f5.6 auto-ISO @ 10000)

Third, down the road when DxO Deep Prime became available (before XD) — looking at the date here I guess I processed this in 2020 — I tackled the RAW again and simply fell out of my chair with how good the image looked. Sure, there's some smoothing in the skin and sofa, but look the neatness and detailed in the hair, eyebrows and even shirt fabric, while keeping the shadow noise reasonable and hanging onto colors — I mean, it just looks like you took this picture with a completely different camera, or at least a couple stops ISO lower than the above pictures.

It's only a tiny exaggeration to say that once I saw what DxO Deep Prime could do with this image, I never worried about noise again. Obviously, there's no substitute for having good light, a lens with a larger aperture, and keeping shutter speeds as low as you can for the subject matter to maximize exposure. And obviously, you're going to be limited in terms of final output size & quality compared to lower ISO output... But if/when you need to really push the envelope, it's really remarkable what today's software magic can do.


GX85 RAW processed in DxO PhotoLab with DxO Deep Prime in 2020 (1/60s f5.6 auto-ISO @ 10000)

• Suggestions for the types of subjects that I can easily shoot myself that respond very well to AI noise reduction. Pets, flowers, hand held still life? I have lots of champagne bottles. :-D

I am also going to cover causes of noise, techniques to minimise noise in camera, e.g. ETTR, HDR, Image stacking etc., traditional noise reduction before AI techniques, and the benefits to your photography of not having to worry about your ISO settings. I will give some examples of how to use Denoise in Lightroom and I will cover software costs, file sizes, processing times, etc. Does anybody have suggestions of other areas I should cover?

I think it's a super-obvious thing, but in addition to the many great messages you're going to deliver, one you'll also want folks to leave with is this:

If you want to get the most out of your images with AI noise reduction, always shoot RAW.

Even if you want immediately JPEGs, too, shoot RAW+JPEG so you'll always have the ability to go back to the RAWs in the future as photo processing software improves.

By the way, this sounds like a really amazing talk. Maybe you can record it and upload to YouTube and/or share your slides?

PS: If you like the example I provided, feel free to use in your talk. I can also send you the RAW file if you like.

 

[Digital Nigel]

One point about NR is that it's really easy if you don't care about details. It's just controlled blurring, after all. So, older NR tech just smudged away the details along with the unwanted noise. It's what happens in-camera.

So the tricky thing is to remove the noise while retaining the fine detail. Either detail gets lost, or fake details get created (because it's very difficult to distinguish between faintly discernible real details and mirages in the clouds of noise). That's where the AI comes in. It's much better to apply it to the raw file before any other processing has disrupted the details. But that can be a problem if the NR tech isn't built into a competent raw processor.

 

[Chris R-UK]

I probably would not dwell on processing times, as that is so hardware dependent. In my case, Lightroom AI Denoise is actually faster than Topaz Photo AI. In other extreme cases seen on this forum, members with different hardware report LR as an order of magnitude slower than Topaz.

My hardware is not high end: $1200 Dell G15 with i7-12700H and RTX3060L

I was just going to quote times with my 20MP Olympus files on my desktop, and warn them they will get long processing times with old computers and high resolution files.

I find using DXO that batch processing 40 files is very much faster per file than processing single files because DXO works on 5-6 files simultaneously.

--
Chris R

 

[Chris R-UK]

Good point about raw. I will be sure to make it in the talk.

 

[Chris R-UK]

One point about NR is that it's really easy if you don't care about details. It's just controlled blurring, after all. So, older NR tech just smudged away the details along with the unwanted noise. It's what happens in-camera.

So the tricky thing is to remove the noise while retaining the fine detail. Either detail gets lost, or fake details get created (because it's very difficult to distinguish between faintly discernible real details and mirages in the clouds of noise). That's where the AI comes in. It's much better to apply it to the raw file before any other processing has disrupted the details. But that can be a problem if the NR tech isn't built into a competent raw processor.

That is a very useful explanation. I will use it.

Thank you.

 

[Digital Nigel]

Will ths be an in-person talk, using PowerPoint slides? If so, you need to take into account the very low resolution and possibly poor sharpness of the projected images. Noise and details will only be visible if you show small crops from the images.

You probably need to find some images where the aggressive AI NR's detail recovery has produced artefacts. At normal viewing distances, these fake details aren't usually a problem (they can even enhance the image), but they can be very ugly if you pixel-peep.

Artefacts can be a particular problem with human faces, because our brains are better than the computer at spotting false details in faces. We evolved to instantly spot and recognise human faces, and know what mouths, noses and eyes should look like, even with complete strangers. They're much less of a problem in feathers, fur and foliage, as we don't usually spot the fakes. So you should turn down the detail recovery with faces, but leave it high with nature scenes.

 

[Chris R-UK]

Will ths be an in-person talk, using PowerPoint slides? If so, you need to take into account the very low resolution and possibly poor sharpness of the projected images. Noise and details will only be visible if you show small crops from the images.

You probably need to find some images where the aggressive AI NR's detail recovery has produced artefacts. At normal viewing distances, these fake details aren't usually a problem (they can even enhance the image), but they can be very ugly if you pixel-peep.

Artefacts can be a particular problem with human faces, because our brains are better than the computer at spotting false details in faces. We evolved to instantly spot and recognise human faces, and know what mouths, noses and eyes should look like, even with complete strangers. They're much less of a problem in feathers, fur and foliage, as we don't usually spot the fakes. So you should turn down the detail recovery with faces, but leave it high with nature scenes.

Yes, it will be an in person PowerPoint talk. I will be showing 100% crops.

With my soccer images I have had particularly problems with AI NR and.faces, probably made worse by the floodlighting. If I turn the NR up, the faces lose almost all detail and they look plasticky. At lower NR levels there are artefacts/noise in the faces at 100% viewing, but the faces have more structure and texture and look more natural at normal viewing sizes. A bit of noise in the faces looks like beard stubble - in fact it might be beard stubble! There are similar problems with the players' shirts. I will use these as examples.

Where AI NR works really well in my soccer shots is the goal netting with its regular pattern. The ball also cleans up pretty well which is important because the main reason that I am shooting at high ISO is that I need 1/1000 shutter speeds to avoid motion blur. Early versions of Topaz DeNoise used to introduce fairly nasty swirls into some out of focus backgrounds, especially when the background was a grass bank. It was this problem that caused me to switch to DXO Deep Prime instead.

Thanks for your very useful suggestions.

 

[grsnovi]

Hi Chris,

I just installed Neat Image. I have worried about installing an AI-dependent product because my laptop is 7 or 8 years old and has no real graphics acceleration.

He is an example of what it did with an OM-1 jpg that had been previously developed from ORF using Affinity Photo 2.2.


JPG from RAW with AP noise reduction.


NEAT noise reduction applied to JPG

I'm very happy with what it does to existing jpg's!

- Gary

 

[gscotten]

Hi Chris,

I just installed Neat Image. I have worried about installing an AI-dependent product because my laptop is 7 or 8 years old and has no real graphics acceleration.

He is an example of what it did with an OM-1 jpg that had been previously developed from ORF using Affinity Photo 2.2.


JPG from RAW with AP noise reduction.


NEAT noise reduction applied to JPG

I'm very happy with what it does to existing jpg's!

- Gary

This is what Topaz PhotoAI did to your original JPG. Denoised and sharpened. It wanted to upscale it as well but I unchecked that. View them at 100% to really see the difference.



--
George

 

[mmartel]

... Unfortunately for the OP, however, it's not available on Linux.

 

[grsnovi]

Thanks George!

FWIW, I didn't spend $200 and I'm not certain my hardware will handle Topaz AI.

 

[gscotten]

Thanks George!

FWIW, I didn't spend $200 and I'm not certain my hardware will handle Topaz AI.

I got to the Topaz party early and had spent about $120 on their other apps, so they offered Photo AI to me for free. But I understand your point. It is a big chunk of money, especially now that Lightroom has excellent AI noise reduction built in. And doubly so because I ended up paying $1200 for a fast laptop with a decent GPU to get the AI processing times down to a level where it is reasonable to use on a regular basis.

I have a low-end camera and spend a lot of my time in low light venues where I am often shooting ISO 12800 and above, so decent NR is quite welcome.