Equipment

Do Recent Advancements in Upscaling Make the Megapixel Race Obsolete?

By
Published October 19, 2023

If you’ve been in the photography world for any amount of time over the last 15 years, you’ve probably heard the term ‘megapixel race’. Starting back in the early 2000s with the dawn of digital photography, the megapixel race was the race in which each manufacturer of cameras, from point-and-shoots, to DSLRs, to smartphones, competed by showing off how many megapixels each of their cameras offered.

And for the start of these races, megapixel count was considered the most important metric of a camera. With each marginal increase in resolution with new generations of cameras, the ‘Nikon vs. Canon’ debates would reignite, and Camera A could ultimately be considered better solely based on the resolution it could shoot. While I was once a soldier in this fight, I’ve long lost my will in the fight, so for all I know, these silly internet debates could still be going on. But let’s take this opportunity to ask the question, with AI, and advancements in upscaling technologies, does resolution matter anymore?

What Do Megapixels Mean Though?

Megapixel is a metric that measures the overall resolution potential of a digital camera. Meaning literally ‘1 million pixels’, megapixels are a unit of measurement that works with dots per inch (or DPI) to determine how large the images are that come from the camera. An individual pixel is a single square of color information – but when lined up in large squares, these pixels form a mosaic, which then makes your image.

Because of the law of diminishing returns, megapixels used to be inherently more important for image quality than they are now. As you can imagine, the jump from 3 megapixels to 5 megapixels is a 66% increase in resolution from one to the next, and so the image quality would inherently increase as well, simply by the added details captured. But the jump from 24-megapixels to 26-megapixels is far less significant by comparison.

Example of the law of diminishing return

AI And Upscaling Advancements

In recent years, upscaling has gotten significantly better than it was before, thanks to several advancements – from machine learning to the high detail that modern lenses can capture. Upscaling is the process of increasing the size of an image, which usually means you’ll be stretching the pixels to fit a larger space than they were originally intended for. In the past, if you were to upscale an image, and for example, double the resolution size of the image, you’d often get a softness from the image. This is because all you’re doing is stretching those pixels to fit a space larger than they were intended to take, meaning anything lacking sharpness will have those areas compounded, giving you more softness, and a loss of detail. But in the last few years, machine learning and other tools have made upscaling far more successful, as these pieces of software analyze the color data, and upscale the image in new ways to retain better sharpness. While these pieces of software don’t promise that you can take a 3-megapixel image from an archive drive, and turn it into a 24-megapixel image from a modern standard, they do promise to help you save some images that might be unusable based on their lower resolution.

Testing Upscaling

So I decided to test these upscalers to see how good they’ve become in recent years. Those who might follow my work, know that I shoot predominately in studio, using the Fuji GFX 100s – a 102-megapixel medium format system that could be regarded as the current winner of the megapixel race. So to test these upscalers, I’m going to take some of my full-resolution images, downscale them to be roughly 30 megapixels, and then use the individual upscalers to bring them back up to 100 megapixels to see which one does the best job.

For this test, I’ll be using a couple of different pieces of software – Capture One basic upscaling, Adobe Photoshop 2023, and Topaz Labs Gigapixel AI and comparing them to the original image. Because I cannot show 100-megapixel images on this blog (for loading reasons), I’ll instead opt for crops of each image to give the best representation of each piece of software. But first, let’s look at the tools and techniques I’ll be using for each upscale.

Capture One Pro

Unlike the other two pieces of software that have machine learning built into their software to handle upscaling, Capture One Pro seems to focus on the old-fashioned way of upscaling – stretching the pixels. For that reason, Capture One Pro will largely be used as a control for this test. For their software, I’ll simply be taking the downscaled image, importing it into Capture One, and then exporting it with the resolution increase set.

Adobe Photoshop 2023

Adobe has shown off some incredible AI tools within their latest release of Photoshop – one of those recent updates comes with their “Super Resolution” tool, which promises a 4x increase in resolution for photos. So for that reason, I’ll be using that toolset, which is traditionally found in the ‘Enhance’ menu of the RAW image importer.

Topaz Labs Gigapixel AI

Finally, I’ll be testing Gigapixel from Topaz Labs – which seems to be the favorite from forum and Reddit posts. Topaz Labs makes use of AI to increase the resolution of images using a couple of different methods available depending on what you’re looking to have done. I’ll be using the most practical methods based on the images tested, which is ironically called High Res.

The Tests

I took four images from my portfolio, downscaled all the images to roughly 30 megapixels, and then upscaled them back to their original 100-megapixel size, below are the results of these tests. Clicking on the comparison image should open them up in full resolution.

Image 1


Image 2

Image 3

Image 4

Results

Given the massive resolution of the images, I had to take tight crops of each photo to make the comparison. I looked for places on the photos that contained the smallest details, as it’s usually where upscaling struggles the most. Overall, I was pretty impressed with all of the upscaling options available, but it seems that Gigapixel did far and wide the best overall. In particular, Gigapixel did the best in those micro contrasts in the images (though it admittedly struggles a little in color options (the lashes on Image 2 are a good example).

But to go back to the question in the title of the article – is the megapixel wars finally over? Yes, I think so. As long as you have a camera that can shoot 25-30MP, you should easily be able to upscale the images into larger formats without anyone noticing. The introduction of machine learning over the last few years has made upscaling better than ever before, and as such, your megapixel count is less important than ever.

But which upscaling option was your favorite? Feel free to chime in in the comments below.

Related Reading

Author: Zach Sutton

I’m Zach and I’m the editor and a frequent writer here at Lensrentals.com. I’m also a commercial beauty photographer in Los Angeles, CA, and offer educational workshops on photography and lighting all over North America.
Twitter Facebook
Tweet
Posted in Equipment
  • John Gaylord

    This test doesn't show what happens to lower-res images containing artifacts such as moiré, which could be present in fabrics and other uniformly patterned materials. The color and gradation contamination of moiré would undoubtedly be magnified without being corrected. In my experience, there is no substitute for either a high res sensor or else multi-shot merging.

  • Mike Earussi

    What I would be more interested in seeing is a shot with your 100mp camera vs a 24mp one that had been upsized to 100mp.

  • While starting with a 24 MP image vs the 100 MP GFX 100 files would be technically an interesting proposition. Zach’s test shows a true result. Gigapixel’s extraordinary upscaling is the best of the three main ways to upscale and image and in most cases, it’s good enough to triple an image. This is significant for those of us shooting stills on an A7S III or a ZV-E1. Our stills are of very high quality but modest resolution. If for some reason, one of us wanted to make wall size prints of our work, Gigapixel would do the trick in most cases.

    The resolution war is over. Announced by a guy who shoots 100 MP images. If such a photographer is willing to admit the resolution war is over, why aren’t the rest of us willing to let it go?

  • User Colin

    As other’s have noted, downsizing to 30mp isn’t the same as having a 30mp camera. The solution would appear to be to take a Sony A7S, A7 and A7R and shoot the same thing with the same lens and very careful focussing. Then see if you can turn the 12.9MP into 24MP or 24Mp into 61MP.

    Even without AI, I did tests comparing downsizing by 50% and upsizing back again with standard algorithms along with a wee bit sharpening, and it is hard to spot the difference even when pixel peeping. Essentially, all I saw was the noise dots in the sky jumped around but the actual real detail didn’t change much.

  • Alex Greenfield

    Technically, this would be possible but would require an ASIC specific implementation of an AI upscaling routine to work in the power and time constraints imposed by a portable camera with a battery and a tight power envelope.

    I see no benefit to it if we are already editing our RAWs in post and now have 1-click hardware accelerated upscaling at our fingerprints. Phones are another story, however.

  • Eric Bowles

    I wonder how long it will be before the camera companies incorporate upsizing when the file is rendered. I think we are seeing that on smartphones now. Take a 24 megapixel full frame image and upsize it in the camera using software so it is produced as a 96 megapixel file. With AI based software, if it’s possible in post it should be possible in the camera.

  • ShullBit

    Impressive results, but you might want to try other subjects, not just portrait photos. AI upscaling depends on the software’s understanding of the photo content to create the right patterns. Therefore, results may vary for landscapes, pets/wildlife, architecture, food, fashion, plants and so on. I tried Topaz Gigapixel a year ago with some wildlife portrait shots and was not impressed. Details were not better than with very old bicubic upscaling and Gigapixel added some weird patterns. Maybe Gigapixel is better now. Maybe it’s just not as good on wildlife subjects like on human portraits?

  • GaryW

    You guys are technically correct that downsizing will give a better result than a 30mp “native” image, but the point of the article was to give a fair comparison of the upscalers. It’s going to be hard to upsize a 24mp image and know for sure if some detail is accurate or not. However, I think it should be possible to get close, by using a zoom, and having a zoomed-in image to compare. Still might be trickier to do the comparison compared to what Zach has done. I doubt the color depth difference (between the downsized and native photos) will affect the results of the resolution comparison much if at all. Maybe Zach will be bored enough one day to make another test. Regardless, all of the programs do a really great job of upscaling, and it starts to fall under diminishing returns. In many cases (landscape, perhaps?) it may not matter if false data creeps in, as long as it looks nice and sharp.

  • Franck Mée

    There’s a caveat here: it seems you didn’t take into account the impact of the color matrix. Basically, when you divide the pixel count of a full, demosaiced image by four (and 150?30 MP is quite close), you put a whole Bayer quad into one single pixel. So you get the spatial resolution of 30 MP, but you also keep the color resolution of ~120 MP. When blowing images back up, algorithms have to improve spatial resolution but not so much color. Therefore, it’s not the same as shooting a 30 MP sensor (which has ~7.5 MP blue and as many red) and blowing it up to a full-color 150 MP picture.

  • Steven Kornreich

    Question.. doesn’t down sampling actually increase perceived detail of the original 100mp image then upscaling it give a better result.
    I maybe totally wrong here btw.
    What I am questioning is if we took the exact same image with both a 100mp camera and a 24-30mp camera the 100mp camera by default will always capture more detail. The upscaling process as good as it is today is adding perceived detail utilizing AI

Follow on Feedly