I’ve tested PhotoLab 9.4 noise reduction. You can see the results! 
AI = Artificial Imagination ?
1 Like
All noise reduction gives artifacts. Even classic noise reductions algorithms can confuse noise for signal. You have to pick your poison.
What do other noise reduction methods look like for the sample you provided?
1 Like
Remember when TopaZ added Ryan Gosling faces to noisy images? Gigapixel AI Accidentally Added Ryan Gosling's Face to This Photo | PetaPixel
Still one of my favorite moments on the internet.
3 Likes
My son picked an AI image the other day because it had six fingers…Hmmmm…computers were first made to count but AI in this case lost that basic ability.
1 Like
Maybe we use different terminologies, but I have never seen “classic” NR (=pretty much everything without ML) produce artifacts like this. They can’t, as they are not generative. Oversmoothing is the most common problem, and I occasionally see color casts, but those can be mitigated.
Classic NR algorithms are so robust that cameras include them to run on their tiny processors. They are a mature technology.
2 Likes
artifacts aren’t just an issue with ai: see Xerox scanners/photocopiers randomly alter numbers in scanned documents [D. Kriesel]
3 Likes
I was in the audience for that one. That was such a fun talk!
1 Like
Sure, AI is not perfect, and artifacts simply exist not only in AI algorithms.
But, when I take a look at the expectations and the amount of money spent on development, there is a huge difference.
Will AI ever meet our expectations about “intelligence”?
AI noise reduction pretty much always produces strange artifacts when there are high-level details at a very small scale.
Personally, I think it’s overrated as the vast majority of photographic situations don’t require such advanced techniques if you pay the basic attention to your camera settings. And it’s also rather disrespectful to our limited resources, using megawatts of power to endlessly train models so we can get just a few more photographs with lower apparent noise levels. Way past the point of diminishing returns for any real value beyond decadent amusement.
9 Likes
In the pre-Christian Hungarian belief systems, people with born with six fingers (which happens occasionally) were assumed to have supernatural powers and usually ended up as a kind of shaman. So it may be that AI is trying to say something here, nudging us in a spiritual direction.
Based on the history of AI, that depends on where you are in the hype cycle.
- beginning: we are creating godlike superintelligence any minute now
- pre-peak: we are a few years away, give us funding
- post-peak: you are asking the wrong question, don’t be ridiculous, that was never a goal
2 Likes
i generally agree with the sentiment. just to put this into perspective: for simple neural denoising (cnn/u-net, say) you can train the model on your desktop gpu in a day. the light oidn model has just above 300k parameters, that’s way less than a megabyte. while it may be inelegant, uninspiring, and incredibly bloaty as compared to conventional methods, it still executes in some 10s-100s of milliseconds (depending on many things such as hardware and resolution of course) and is certainly not the class of algorithms people build nuklear reactors for.
also: smaller models: less space for hallucinations.
3 Likes
Over the last year, I got more into printing. Funny thing about prints: noise just looks like texture. It does not detract from the subject. In fact, in many cases, it increases the sense of sharpness and presence of the picture.
I have since largely stopped worrying about noise. It only looks bad with nearest-neighbor downscaling, but that’s a failure of the image viewer, not of the image.
4 Likes
That is certainly a good point. Except for the most ludicrously noisy situations in which case I think it’s better to try something else and quit taking photographs, a little noise isn’t visible at all in prints. Yes, nearest neighbour does look quite horrendous even with the most mild of noise indeed.
I truly don’t understand the obsession about noise removal. I have seen MANY beginners praise the software only to then show a bunch of their shots before and after where the before hardly had any noise to begin with. I think we have been trained to focus too much on technical solutions, forgetting that there is not even any problem to solve.
Indeed, that’s true. There are some models that are very small.
3 Likes
That’s what I dislike about sending my photos to family over WhatsApp - it looks A LOT noisier that the original. I guess I gotta downscale first and hope the resolution is low enough
1 Like
The same applies to sharpening as well, regrettably. Pixel level sharpness is entirely invisible in the printed picture.
Really, sharpening only ever makes sense with respect to the output size. But that’s bad news, as I want to be able to look at my photos both on the phone and on the computer screen.
At least for printing, the printer driver usually comes with a decent built-in sharpener that fits to the print size. For whatever reason, Canon calls this “Use Contrast Reproduction” 
.
To tie this back to the thread, this applies to AI sharpening, too.
Are we sure that’s not what things looked like? Maybe the AI just sees through the holographic image tech all the lizid people wear… 
1 Like
That is true, which is why capture sharpening for example, while a nice algorithm, doesn’t really make sense for a lot of photos unless you crop them a lot or are displaying them at a super large size. One must select the sharpening for the output. I honestly think if people printed more, there’d be a lot less confusion about all this stuff!
2 Likes
Frequency domain noise reduction can give false patterns that are something like JPG artifacts. Block matching noise reduction can make noisy straight edges into a jagged line.
All algorithms have artifacts (except maybe the exposure module, but someone will correct me ). But from my perspective, a jagged line I notice while pixel peeping is a far cry from an extra hand. YMMV.
2 Likes
My photos are meant to stay digital, being watched on a computer monitor, very often zoomed in and out, so low ISO capturing and fine, gentle sharpening of details is a part of full-size hi-quality export.
On my walls, I have paintings, some with many tiny details, that require watching from various distances, no one is the right one. 