[This post first appeared in Photo AI, my newsletter about how computational photography is fundamentally changing photography as we know it. Subscribe here, or view the archives at Photo-AI.com.]
I’m going to be a contrarian today, in a specifically picky way. You may have seen that DxO this week released DxO PureRAW 3, the latest version of its pre-processing software that uses AI to improve raw files. If your browsing is like mine, you’ve also seen numerous articles and videos praising the new DeepPRIME XD technology that removes digital noise and improves detail recovery.
Don’t worry, I’m not here to trash it! I wrote about DxO PureRAW 2 last year, and continue to use and recommend it. From what I’ve seen in my limited testing of version 3, the update looks great. Absolutely check it out; you can use it in an unlimited trial for 30 days.
But I’m also struck by some of my results and how the cult of noise reduction might push some people toward a too-soft future.
But first, let’s back up and look at why you’d want to interject PureRAW into your workflow in the first place. Remember that a raw file isn’t actually a photo yet: it must first be decoded and demosaiced in order to translate the raw data into pixels that can then be edited. This is why, for example, a raw file must first pass through Adobe Camera Raw when you open it in Photoshop. The different companies each use their own decoding software, which is how you end up with Capture One producing a better Fujifilm .RAF raw file than Lightroom.
DxO PureRAW sits at that first stage, using its expertise in raw decoding and vast amounts of data covering nearly every camera sensor and lens characteristic. I shoot with Fujifilm cameras that use Fuji’s X-Trans sensors, which don’t always render well in some apps such as Lightroom, so I often run raw files through DxO PureRAW to get better results.
Part of the demosaicing process is dealing with digital noise. Since pixels are not yet burned into the image at this stage, raw files include data about the characteristics of the camera sensor and how to deal with it. DxO PureRAW uses machine learning to evaluate and correct for digital noise, which can be amazingly effective in images that were captured in low light situations or at high ISO settings (which are often combined).
Let’s put digital noise into perspective. The sensors in today’s cameras are much better at dealing with noise than in the past. We used to caution photographers not to increase their ISO too high; approaching ISO 1600 was like taunting the sky dragons when test pilots bumped up against Mach 1. You could do it when you needed to increase the sensor’s light sensitivity in dark scenes, but you risked too much unwanted digital noise.
Now, ISO 1600 is barely noticeable on some cameras, and it’s practical to shoot at ISO 6400 or higher. When you do need to break into those higher areas—which do create noise, don’t get me wrong—tools like DxO PureRaw or ON1 NoNoise AI can clean up the images.
So I tossed a few problematic photos at DxO PureRAW 3. This image was shot using a Fujifilm X100V during the very last light of the day at ISO 6400 (left). The PureRAW 3 result, using its new DeepPRIME XD technology, is at right.
No doubt, the software has eliminated the noise, but at 100% it’s too aggressive for my taste. When zoomed out, the effect is better, but it still exhibits that pastel-looking smear emblematic of AI processing (like many low-light smartphone images).
I ran the same image through DxO PureRAW 2, which uses the DeepPRIME (not XD) technology. (You can choose to use DeepPRIME as the processing option in PureRAW 3; I just used version 2 for this example.) To my eye, that’s a better result, particularly compared to the XD version; it still has noise, but it comes across more as grain.
One more noisy example. A few years ago I photographed a lantern-lighting ceremony at Green Lake here in Seattle with my Fujifilm X-T1 camera, which has a 16 MP sensor that didn’t boast great low-light sensitivity. I didn’t want to blast people with a strobe flash, of course, so I pushed my ISO to 4000 and hoped for the best. A lot of the shots were… not great. Here’s one example, straight out of camera with no processing applied.
When I click Auto in Lightroom’s Develop module, though, I’m reminded how much light data is collected in a raw image even in a nearly-10-year-old camera. There’s nice illumination on the woman’s face, which wasn’t apparent at all in the unedited version.
Zooming in, however, reveals the ISO penalty I paid. So I processed the file using PureRAW 3 and then clicked Auto.
Again, the noise is gone, but DeepPRIME XD has considerably softened the face and added a halo under the woman’s chin.
New in this version is a setting for dealing with lens softness. I reprocessed the raw file using the Hard setting (the most compensation), which added contrast and removed the halo.
I think that goes too far (and adds some sharp artifacts to the people in the background), but it’s nice to have some flexibility.
(By the way, one tip for bringing the illusion of detail back to de-noised photos: add grain.)
Going back to the other part of the PureRAW appeal, better raw processing, I’m happy to see improvements in DeepPRIME XD for rendering details in my Fujifilm photos. Capturing distant autumn leaves is sometimes frustrating because Lightroom tends to make them look like brush strokes. DxO PureRAW 2 was a revelation for me, and version 3 does an even better job.
Granted, those noisy photos are two admittedly extreme examples. My contrarian point is that photographers might do what I did and just throw DeepPRIME XD at everything, happy that their noisy images are cleaner and brighter, but won’t consider that they’ve incorporated the kind of mushy smoothness that AI de-noising is supposed to avoid.
I’d hate to see us in a few years looking at these images the way we now look back at early HDR experiments with their garish saturation and over-sharpened everything.
But then we’ll still be able to reprocess the original raws, so that’s good.