Menu
#1306. AI upsampling is here. Can I have my large pixels back, now?
By pascaljappy | Opinion
Can we all agree that 100Mp is enough for anyone? And that larger pixels are better pixels?
Have you ever compared modern, high-resolution files at 100% to older, lower-resolution, photos at 100%? The comparison is not always flattering for the more recent offerings.
You might question the purpose of looking at a file at 100%, but my reply would question the goal of high-resolution cameras if all that resolution isn’t really used.
Because, that’s a real issue. It’s extremely rare that the theoretical resolution on offer is actually fully exploited by the lens/technique combo producing the file. Particularly with the higher density sensors in recent hi res cameras.
Let’s look at the advantage of lower-resolution files for a minute :
All of these are time/money savers and significantly better for the environment. Plus lower res is kinder to older lenses, which tend to have more charm than newer ones, and less stringent on AF precision, data throughput in the camera.
Now, what are the benefits of higher resolution?
Some users of 24Mp FF cameras swear it is the ideal compromise. Not ever having owned such a camera, I cannot comment on this. But the mesmerizing quality of *some* Leica SL2-S photographs, in the hands of excellent photographers, do lend this theory some credibility. Those files look more film-like than anything else I’ve seen produced by a CMOS machine. And the recent trend of offering lower res files in the latest high-res machines is noteworthy as well (although it seems those lower resolution modes fail to deliver the promise of higher quality, presumably because assembling tiny pixels ain’t the same as shooting large pixels from the start).
Yet, for the rest of us, nursed on high-count megapixel digital files, the 24Mp turkey might be too cold to handle. Withdrawal symptoms could rob us of the joy of shooting, at least initially.
Here’s where upsampling comes in.
No, upsampling 24 to 48 is not the same as capturing 48 in the first place. Lost information cannot be invented in PP. But this brings us back to the initial question: have you compared 24Mp and 48Mp files at 100%? Upsampling from a tack sharp image might not be far behind the slightly mushy high-res image. Just sayin’ …
Besides, a whole wagon of new upsampling apps, powered by AI, is now available to us. This article alone lists 7 beyond those included in software such as Lightroom.
Here’s the kicker: each of these has its own look. Meaning you can incorporate upsampling in your creative process. It won’t have as big an impact as choosing a different lens design, but it’s not to be sniffed at.
Just for fun, here’s what the neutral enhancer in Lightroom does on one of my X1D images. The first image is the whole frame. The second is a 200% crop of the enhanced (equivalent to 200 Mp) image. Is that significantly worse than what your standard files look like at 100%? A 24Mp file upsampled to 100Mp would look as good.
There’s been this downsizing trend in architecture, and other areas of our lives, recognizing – finally – the difference between quantity and quality. More is not always better.
Now that the megapixel race seems over, how about a race back to the future, offering cameras with all modern trimmings, all the bells and whistles (IBIS, AF, ISO …) but not the nefarious resolution?
I’m looking at you, Hasselblad. How about an X3D with the fantastic IBIS of the X2D, all the cool modes, the improved colour science and … 40Mp, if you can find such a sensor in someone’s parts bin? Higher DR, better AF, smaller files, smaller buffer, the best tech combined to larger, much laaaarger pixels! Sony ain’t the only sensor maker out there …
Here’s hoping for quality to finally get its revenge on quantity. I’m not saying all cameras should be lower res. Only that not all of them should be hi res. Give us a choice.
Never miss a post
Like what you are reading? Subscribe below and receive all posts in your inbox as they are published. Join the conversation with thousands of other creative photographers.
#1307. Reclaiming the Elbow
#1305. My view on generative AI in photography
#1304. Leica SL2 and Nikon 50mm AIS Review
#1291. Slightly Off Topic: AI and photography blogs
#1302. Objective vs Subjective photography. Is there a middle ground ?
#1290. Three more thoughts about (experiential) photography
Session expired
Please log in again. The login page will open in a new tab. After logging in you can close it and return to this page.
Wonderful post, especially relevant to something I am wondering about on a photo I may submit as part of the August collection (https://www.flickr.com/photos/159837371@N07/53099500951/in/dateposted-public/). While on holiday I had only my GR iiix and wanted to make an image of this house on Eastern Harbor, Maine. Intentionally made using the 71mm in camera crop, reducing the image to 7mp. Further cropped in post processing, then upscaled in Gigapixel AI. I like the image, but am I cheating? It is an extreme violation of the recommendation (rule?) not to crop!
Thank you Dave.
There’s nothing wrong with cropping if it’s not an alternative to moving with your feet. Just like using a zoom in a lazy manner. Getting closer changes the perspective and the composition with respect to using a long lens or cropping. Apart from that, I see no reason not to crop 🙂
Cheers
Hear, Hear!
Btw., a question:
For economical reasons I guess I’ll stay with APS-C (and avoid the rather expensive “modern” lenses).
Pascal,
do you have a suggestion for a suitable number of pixels?
( – from your quality standpoint.)
Just curious, such a sensor will probably not be available (soon) for mirrorless…
Cheers!
Hi Kristian, I don’t have a standard quality in mind. But if a large number of FF users like 24Mp, a similar pixel size would yield about 10Mp. But I remembe liking my 16Mp APS-C Sony NEX a lot 🙂
Cheers
I love my D500, Kristian, which I think has 20MP.
You won’t get any argument out of me, Pascal.
I decided several years ago that this whole argument was just silly. AFTER I bought my D850 and plugged my Otus 55mm into it, I decided that was fine for some stuff – buit unsuitable for things like pet & animal/bird photography, so I bought the older HF D500 for that – and it works just fine – with half the pixels, and a smaller sensor.
Then I started to see article after article by professional photographers, telling me that for the most part something around 23MP was tones enough. “Some users of 24Mp FF cameras swear it is the ideal compromise.” Quite – over & oveer!
While I was absorbing all of that, I was making larger prints of my own stuff. One theory is you don’t need more pixels unless you DO make larger prints. But even making larger prints didn’treveal anything deplorable. Let’s face it – you don’t look at a larger image from 10 cm away – the larger the image, the longer the viewing distance, and the less likely you are to worry about pixel counting. Of course I’d always known that, from film days – who cares about “grain” in the image, when you stand 2 metres back, to look at it? Why should pixels be any different? And of course, they’re not.
“Can we all agree that 100Mp is enough for anyone? And that larger pixels are better pixels?” Yes and no. Or rather no and yes. Because actually I think 100Mp is way more than necessary. But I do think larger pixels are better than smaller ones. Because like any other container, they can hold more information than smaller ones. If this was a mathematical issue, a rectangle covered in circles would, I imagine, have much the same surface area “within the circles” as “outside the circles”, regardless of the diameter of the circles. Dunno – just imagining it. But that’s not where this ends. Time after time I’ve seen statement after statement by photographers who know vastly more than I ever will, saying “bigger is better, because they hold more information”.
Bottom line – all I see in tis argument is that for most purposes 23mp is quite sufficient – but sensor size gives better dynamic range than smaller sensors can – which might be just an illustration of the advantages of larger pixels over smaller ones. Who am I to say? All I really know is that I long ago lost interest in this argument, and I can’t see myself ever worrying about it again. Anywhere between 20 & 45 suits me perfectly. And when Canon tried to upend Nikon’s success with its 45MP D850 by introducing a 54MP sensor, all they succeeded in doing was to give their customers a visibly poorer image quality.
PS – I must do something about my typos – I can spell better than that, so it must be my keyboard skills! Sorry, folks.
Hey, no biggie, Pete! Sorry I didn’t catch the typo (still don’t see it), I would have edited out. But it’s nothing to worry about.
Pascal, some lovely images at whatever MP, for me 45MP of the Z7 is good for my eyes.
Thank you Dallas 🙂
Your recent photos sure speak highly of the Z7! Though I believe the photographer holds more reponsibility in their success 😉
Hi Pascal
I have a couple of comments.
First, more pixels give us more color data, which should be better. With a higher resolution sensor there are more pixels covering any given area in the scene, so your camera or processing software should be able to calculate the color variations over the area more precisely. It would also benefit us if our camera, or processing software, would give us a choice of interpolating color over 9 or 16 pixels vs. 4 pixels.
Second, I haven’t really “compared”, but I have studied my results from 20.2 MP (M43) images and 47MP (FF), and I can honestly say that 20.2MP processed in Capture One is quite pleasing.
The thing I have noticed in doing enlargement comparisons on a 4K monitor is that going to 100% does not give me the satisfaction it did with older monitors and graphics cards. My 20.2MP images look great up to about 67%, but go soft fast as I increase the zoom factor. With 80+% being about the limit. I notice a similar effect with my 47MP images looking great up to 75-80% then being too soft around 115-120+%. These effects vary some with subject and distance to the subject. Oh, and my technique capturing the images.
Considering your print crop suggestion, if memory serves me well (which it may not), I would say your enlargement of WON 220 is fairly close to my 20.2 MP screen image at 80% and 220dpi. Of course, the proof of the print, is in the print. Our screen views just don’t compare.
PaulB
Hi Paul, up to a point, I would agree with you.
But as pixels get smaller, the signal to noise ratio degrades. Some noise is independant of pixel size. But the number of photons you can catch grows with the surface of the pixels. So a 4 micron pixel will catch 4 times as many pixels as a 2 micron pixel, for example. That’s a big difference in potential dynamic range and signal/noise ratio.
A higher S/N ratio means much cleaner data. And, with the quality of today’s upsamplers, I do believe you’re far better off starting from very high quality data and upsampling when you need it than using small pixels and degraded data. More pixels also mean more amplifiers, smaller circuits, more noise …
The M43 sensor is tiny. So 20Mp actually means very small pixels, which will not perform as well as the larger ones in a 20Mp FF camera, for instance. Of course, there are other advantages to the M43 system, such as incredible IBIS, to name just one. I’m not dissing smaller sensors. Rather, I think that systems with large sensors would benefit from lower resolutions, today.
WON 220 is WON 55 viewed at 220% after AI enhancement of my 50Mp files in Lightroom. It doesn’t upsample, it simply enhances details already here. My hunch is that it would perform better with a system that doesn’t already produce so much detail. Anyway, I just presented the screen grab at 200%, which is equivalent to seeing a 200Mp file at 100%.
The X1D is still my favourite camera out there, simply because the quality of its images has to be seen to be believed. There seems to be zero degradation. Everything is super clean, super sharp, noise look great. And I feel the look from the 100Mp is more clinical and less clean. So my preference goes to the older model 🙂
Cheers
Pascal
I do wish we lived closer to each other, which would make this conversation would be so much better in person over a nice adult beverage. As I think we are approaching this from opposite sides of the same coin.
Your comments about the higher signal to noise ratio from a sensor with larger pixels at one time was an absolute truth. Today, with sensor and digital signal processing improvements, it needs a slight modification, to “It Depends”. Today, when we compare sensors of different physical sizes, you are correct, the larger pixels on the larger sensor will result in better image quality. On the other hand, if we are comparing sensors of the same physical size, the sensor with the larger number of pixels will have better image quality. Since the newer digital signal processors can reduce the noise collected from the pixels to a greater degree than just averaging across pixels would imply. Of course, if we are comparing the Fuji 100 MF cameras to the Sony A7RV, then all bets are off. Because both sensors use the same production technology and both have the same pixel density.
Now, having said the above, Kirk Tuck and Damien Lovegrove have both said that they prefer the image quality from the Fuji 50MP MF sensors over the Fuji 100MP MF sensors. So, I think you are in good company. I certainly envy the results you post here. As I can see a difference between your images and mine, even though I shouldn’t on the web at 1500 pixels.
Concerning my comment about the 20.2MP from my M43 sensor. I wasn’t really commenting about the size of the sensor, but about the number of pixels and that I am pleased with my results. To a large extent once an image is captured pixels are pixels, so the number only matters to a limited extent.
Where, I was saying that more is better is with color data.
For the sake of discussion, let’s say that 100MP on a full frame sensor is the resolution limit that current lens technology can deliver without increasing lens prices exponentially. So, we might think that once sensor resolution reaches 100MP we have hit a performance wall and there is nothing to gain from more pixels on a sensor. This would not be true, if the sensor manufacturers were able to reduce the pixel size by 50%, the number of pixels would increase from 100 to 400. This could be huge benefit to us in terms of color. Since, even though our lenses can only project to a resolution of 100, we would get true color data from the same pixel area. Ideally this could go another step further and the sensor filtration could be changed from Red-Blue-Green (RGB) to Cyan-Yellow-Magenta-White (clear), which might provide an even larger color spectrum to work with.
Circling back to your comment about the images being less clean, you may be right. The higher resolution sensor may be capturing something that 50 MP doesn’t. Much like the comments about digital audio recording when digital music CDs were coming on the market that the technology was too good. And could you tolerate listening to a symphonic recording if you could hear that the piano pedals squeak when the pianist plays.
PaulB
Fascinating!
And the analogy with recorded musing adds a layer. I think initial CD players were terrible, so we couldn’t draw conclusions. But today’s high-end offerings in streamers, CD players and streamers are more or less on par in terms of “performance”, but their sound is very different. This is where the discussion gets tricky.
Like Kirk and Damien, I prefer the look of the 50Mp sensor to the look of the 100Mp sensor. But I loved the look of the larger 100Mp sensor (56x44mm) and even prefer medium format film (photography’s vinyl). All this is preference and personal. And I’m not saying one is inherently superior to another.
Here’s the quantitative difficulty: all things (technology in particulary) being equal, a larger pixel will provide a better S/N ratio than a smaller one. But, given a same sensor surface, there would be fewer large pixels. And the image quality (measured, not subjective) doesn’t depend exclusively on the quality of the individual pixels. You are absolutely correct: manufacturers are able to extract better performance from higher resolution sensors. Certainly, Hasselblad has improved colour science in the X2D over the X1D, in spite of much smaller pixels. But my uninformed guess is that this is through signal management and internal processing. And the result, subjectively, seems less pleasing.
What I would love to see measured is dynamic range. In absolute terms, that is what matters to me most. S/N and DR are the only objective measures I really care about, because they define the information present in the file. With a good file, it’s up to the photographer to eek out the aesthetics and quality they prefer subjectively. So far, my feeling is that this objective envelope is better served by larger pixel sensors, but I may be wrong 🙂
And yes, that discussion would be great over a drink 🙂 🙂
Cheers
What is a court without a jester? But let us not forget that the jester is not merely a clown or buffoon, he speaks the truth. He is even the only who can speak the truth without fearing for his life when he boards a bizjet. Hence my choice to jest. In this spirit, I have to disagree with the ever-clever-and-learned PJ. I see the difference between the more-or-less-standard 24Mp full-size-sensor image, and the 50Mp+ image. I am told that this is impossible, that it is wishful seeing, and so on. Except that the same is true with high fidelity sound. We were told, when going digital, that all components would sound the same, because it was all digital. Then that there was no way that better technical performance would sound better, because the original 16-bit CD white paper standard already exceeded anything the ear would hear. Or, with analog, that no micro-information can be heard when it is smaller than the LP noise level. We all know what to call these assertions. Hogwash! Poppycock! Tommyrot! Now of course, if your purpose is for digital to look like analog, which is low res by dint of film technology, then of course high res. doesn’t matter. Which doesn’t mean that digital is better than anlog in every way, not at all. Merely that higher resolution is visible even on Internet resolution, just as detail below the noise floor is audible on a LP. And if you prefer to bask in the satisfaction of your beliefs than in the pleasure that high resolution brings, namaste! Enjoy!
Hi Philippe, I don’t think it’s wishful thinking. And maybe you can see a difference, like some people can hear a difference. As you know, I streamed for a while a sure heard a difference between 24/92 and Redbook. But I’ve seen Leica M8 prints leave prints from higher resolution cameras for dead. And the difference I’m seeing in high resolution cameras is not one I enjoy.
But, anyway, I’m not suggesting high-resolution cameras be banned. Only that, maybe, there are several tastes and goals out there. And, that some of us should be allowed to buy a modern camera, with all the IBIS/AF/ISO/… benefits, and large pixels.
I’ve now heard too many streamers and DACs for my own good, and the only piece of gear that’s really made me want to reach for my wallet recently was an Ayon CD/DAC that sounded more natural than anything modern I’ve ever heard. And some people, even today prefer vinyl to digital. In HiFi, we have a choice. Why shouldn’t we in photography?
Taste preferences plus all the size/time/energy arguments I mentioned should be taken into account.
Ah, Pascal,
I thought you might suggest somewhere between 10 and 20 Mpx.
And the next gen. half frame sensors will probably be over 30 Mpx…
As to the problem of dynamics, the 12 Mpx Fuji XF1 (and some of the next and previous models) had a nice solution. It had a setting where half the pixels were exposed 1-3 EV less and the two 6 Mpx photos were merged. I’d like a camera with such a 20-40 Mpx sensor.
In the future we’ll instead probably have sensors with continuous readout.
– – * – –
Then there is the square pixel pattern.
We notice lines, and especially vertical and horizontal lines, faster than most other patterns. So wee would be far less sensitive to pixelation with a random arrangement – like film grain – or even a hexagonal pattern.
So, I guess, then our comparison of different pixel densities would turn out differently.
– – –
And there is printing…
Some b/w photos seem to become more appealing on a paper with some structure in it.
( I haven’t seen any comparison in colour.)
On the other hand Ming Thein some years ago on his blog wrote a series of articles on high resolution prints – his printer had developed a process with ink jet that increased the linear resolution by about 50 %.
When looking at large prints of many stitched photos, the most common reaction viewers gave was that it looked more like looking out of the window than like looking at a print.
– – –
So, what do we want with a printed photo?
A pice of art or a representation of some reality – or both?
All very good points, Kristian. The ultraprints were an intereting concept for small walls. A more immersive print to replace a larger one.
I suppose we all want something different from a print, which is fine. And I’m not suggesting there shouldn’t be high-resolution cameras anymore. Only that it would be nice to be offered a choice, at least for those of us who couldn’t care less about resolution 😉
Exactly!
Will we have the replaceable sensor, “…just as easy as changing lenses…”?
And with a (far?) future continuous readout sensor the difference between large pixels and pixel binning ought to be, I think, rather smaller.
Ah, well I don’t know about the continuous readout, but Pixii already offer replaceable sensors. They offer the choice between 3 and are working on other options 🙂
Pascal, Kristian
I’m curious about how each of you are using the term “replaceable” with regards to sensors. Do you mean replaceable as in remove a defective sensor and replace with a new one; though I assume you also mean user replaceable? Or do you mean interchangeable, the way lenses can be changed at will?
Either way, I like the concept. It would be nice if we had the option to remove our color/daylight sensor and and replace it with an IR or B&W sensor. Or, we could send a camera in for sensor repair and have it back in a few days.
In theory it could be done, much like Leica tried to do with the digital back for the R8/9, and the interchangeable module that Ricoh produced. Though, the biggest challenge is creating an electronics package that can be somewhat future proof. So the current sensor module holder (i.e.body) can also accept sensor modules that may be developed a generation or two in the future.
Unfortunately, I think technology needs to progress quite a bit further to make the components small enough and affordable compared to having a second camera body.
PaulB
Hi Paul, it would be great to swap one sensor for another on they fly, wouldn’t it? 🙂
Didn’t a Ricoh camera allow you to do that? You essentially bought a grip with a brain and battery, then added modules that contained both lens and sensor. It was always odd to me that the modules mated lenses to sensors, but there were an interesting form of modularity there that was innovative and fresh.
As far as I know, Pixii can change the sensor in your camera (if you need a global shutter for instance, or if you prefer higher pixel density) but you can’t carry both sensors in a bag and hot-swap them as you would lenses. I’ll suggest it to them 😉 It is a very good idea, actually.
Hasselblad let you change backs on the 907 so maybe there will be a choice of sensors at some point?
Cheers,
Pascal
Pascal
You are correct about the Ricoh camera. It was an interesting concept, but they were too far ahead of their time for it to work. Photo technology changed too fast for them to keep up. If they tried this today, it might actually work. Think of a competitor to the Leica Q or PenF with interchangeable lens/sensor modules. You could have the lens/sensor combinations you prefer and never worry about dust on the sensor.
I also thought of interchangeable digital backs like Phase-One and Hasselblad use. The challenge for this idea, aside from cost, is bulk. Which could make the camera long and thick vs. wide and thin, and that may limit the desirability of the camera to few types of photography.
PaulB
Indeed.
I always imagined that high-resolution sensors used with pixel binning were supposed to produce great results. But that doesn’t seem to be born out by the current 61Mp cameras offering 36 and 18Mp modes. I suppose those are not really pixel binning though. Maybe a 2×2 binned 25Mp readout on a 100Mp sensor would be great???? Hassy, do you hear? 😉
Cheers,
Pascal
Pascal
I think you are correct that the pixel reductions are simply a crop vs binning. A perfect 4:1 binning pattern would give us 15.2 MP from a 61 MP sensor. Which I would expect to provide really nice image quality, because the binned pixels would be delivering true color data for the binned area.
PaulB