X-Trans ex-noise: Testing DxO’s new denoising tech for Fujifilm cameras

This image was captured with a Fujifilm X-T5 at ISO 12,800 and processed using DxO’s DeepPRIME XD denoising algorithm.

It’s been a little over a decade since Fujifilm launched its first camera with an X-Trans sensor, the Fujifilm X-Pro1, way back in January 2012. With a color filter array that samples red, green and blue pixels in every horizontal line and vertical column, X-Trans sensors are designed to more evenly sample the color information across the image. This, in turn, allows better resistance to color moiré versus the more traditional Bayer filter, which skips either the red or blue pixels in every other line or column.

The downsides of the X-Trans approach are twofold. With a larger 6×6 pattern rather than the 2×2 of Bayer, it makes the task of demosaicing the color data not just unique to Fujifilm cameras – requiring software makers to implement more custom code – but also more processor-intensive and challenging to optimize. And French software maker DxO‘s equally unique approach of denoising prior to demosaicing further complicated the situation. So for years it simply eschewed X-Trans support altogether, offering Fujifilm camera support only for its Bayer-based models.

The denoising samples throughout this article were made from these five images, shown here in their out-of-camera JPEG versions. You can click each crop in the rollover galleries below to open the full-sized, DxO-processed JPEG.

In October 2021, that finally started to change with the introduction of DxO’s PhotoLab 5. That brought basic X-Trans support, although a couple of gaps remained in which noise reduction algorithms were available for it. The recent arrival of PhotoLab 6.4 has now closed the biggest of these, bringing support for the company’s top-of-the-line DeepPRIME and DeepPRIME XD denoising algorithms to Fujifilm X-series shooters. (This leaves just the company’s standard PRIME denoising algorithm and a handful of more minor features off the table for your X-series camera.)

So how does DeepPRIME get on with X-Trans imagery in the real world? Let’s roll up our sleeves and take a closer look, courtesy of a selection of high ISO shots from our recent Fujifilm X-series reviews!

All sample images and their crops in this article were converted from raw using the just-released PhotoLab 6.5 update (also available for macOS) with its “DxO Optical Corrections only” preset, and each denoising algorithm was used at its default settings.

Test 1: Bread (Fujifilm X-T4, ISO 2500)

No denoising	HQ denoising	DeepPRIME	DeepPRIME XD

We’ll start off with a shot from our Fujifilm X-T4 review at ISO 2500, a moderately high sensitivity. As you can see, with no denoising applied the noise intrudes quite a bit and, despite the X-Trans color filter array, there’s still a fair bit of false color artifacting on display too.

The base HQ denoising algorithm cleans things up considerably, but it also robs some detail from the finer flour particles stuck to the bread, and the cracking in the crust looks a bit soft at 1:1 resolution.

And while DeepPRIME crispens things up, the newly-available DeepPRIME XD algorithm definitely does a much better job. Its result is even crisper, and it restores more of the fine details, yielding a less plasticky feel than the standard DeepPRIME.

Test 2: Duck (Fujifilm X-S10, ISO 3200)

No denoising	HQ denoising	DeepPRIME	DeepPRIME XD

Increasing the sensitivity a little to ISO 3200 with a shot from our Fujifilm X-S10 review, we have a tightly-cropped shot of a duck’s head with the feathers freshy wetted from a visit beneath the water. Again, we can see quite a bit of noise prior to denoising, especially noticeable in the duck’s eye. False color also intrudes, especially in the reflective water droplets and the fine folds of the duck’s lower eyelid.

Switching to HQ denoising improves both of these defects considerably, but it definitely discards a good bit of fine detail too, throwing the baby out with the bathwater. You can see this most easily in the really fine feathers beneath and immediately behind the eye.

The DeepPRIME algorithm restores much of the detail in these areas, but seen 1:1 it also gives a somewhat unnatural, waxy appearance. It’s noticeable in the feathers to some extent, but especially in the eyeball, which almost appears as if made of glass.

Switching to DeepPRIME XD brings back just a little more detail in the eye and its reflections. It also gives more crispness to the feathers. Again, this result is clearly the best of the group, especially bearing in mind it’s at default settings with no user intervention.

Test 3: Samurai (Fujifilm X-H2S, ISO 5000)

No denoising	HQ denoising	DeepPRIME	DeepPRIME XD

Next up, we have an ISO 5000 shot of a samurai warrior figurine from our Fujifilm X-H2S review. With no denoising the glossy, black panels look quite speckled. And again, there’s a little false color artifacting, which is especially noticeable in the shiny area along the upper edge of the small cord clip which you can see near the top center of the crop.

The HQ denoising algorithm cleans both of these up quite nicely, but it also robs us of the majority of the thread patterns in the fabric at the right of the crop. The threads in the cords are also largely lost, while some unnatural artifacts can be seen in the shiny bronze trim piece at the center of the crop.

The standard DeepPRIME algorithm does a noticeably better job with these defects, but it still looks a little soft when seen 1:1. Once again, DeepPRIME XD looks noticeably crisper, with more of the threads remaining visible in both fabric and cords.

Test 4: Dog (Fujifilm X-H2, ISO 6400)

No denoising	HQ denoising	DeepPRIME	DeepPRIME XD

Our next shot is at ISO 6400 and comes from our Fujifilm X-H2 review. With denoising disabled, it’s quite a snowy affair, especially with the reflections in the dog’s eyeball. Much of the detail in its fur quite heavily obscured by noise, especially in the blue channel. And while the HQ denoising filter helps, it still leaves things quite desaturated.

With the DeepPRIME filter active, not only is that noise mostly squashed, but the colors in the fur and the reflections in the dog’s eye are noticeably more vivid. And that’s even more true of the DeepPRIME XD version. The XD algorithm also rids the fur of some very fine artifacts that are especially noticeable in the top right corner of the standard DeepPRIME version.

Test 5: Street (Fujifilm X-T5, ISO 12800)

No denoising	HQ denoising	DeepPRIME	DeepPRIME XD

Thus far, all of these comparisons have played to the strengths of both DeepPRIME and DeepPRIME XD. As we’ve seen in our past coverage of these technologies – our DxO PhotoLab 6 review, for example – DxO’s AI denoising tends to perform best when presented with naturally-occurring detail. Things like feathers, fur and foliage are its strengths.

Where it can sometimes stumble is with man-made detail, and especially text. So for this ISO 12800 shot from our Fujifilm X-T5 review, we’ve saved the hardest task for last. A nighttime city street scene, it’s not only shot at the highest sensitivity of the bunch but also contains generous quantities of both of these AI bugbears.

As we can see above, DeepPRIME XD continues its winning ways with the ornate moldings at the bottom right of the crop, where your eye is fooled by the crispness and extra detail, and wouldn’t necessarily notice any inaccuracies in the rendering. But these tricks don’t work so well with the shop sign. (And the same holds true of the red advertising hoarding below, if you click through to the full-sized version.)

In both of those areas, the standard DeepPRIME algorithm – while noticeably less sharp than the XD variant – yields more readable-looking text with fewer unsightly artifacts. As humans, we’re programmed to identify things like faces and writing: we can make these out relatively easily even if a bit blurred, but we also notice defects in their rendering more readily.

No denoising	HQ denoising	DeepPRIME	DeepPRIME XD

We’re going to look at a second crop of this image, though, to point out some differences elsewhere within the frame. Here, we’re looking a bit further to the right at the apartment windows, and there are a couple of things to note.

Firstly, DeepPRIME XD seems to be tuned to look for and attempt to replicate patterns more strongly than the standard DeepPRIME. That difference is especially visible in the window blinds; they’re just barely visible in all but the DeepPRIME XD image, which renders them with much greater contrast.

Secondly, look at the roof tiles towards the bottom of the crop. In the DeepPRIME XD version, these have a somewhat unnatural, almost painterly look to them that isn’t present in the DeepPRIME or other versions. This same effect is also noticeable in the background of the sign from the first crop, although there it’s a little harder to tell what’s going on.

Conclusion

From my testing, the arrival of DeepPRIME XD for X-Trans looks to be a win for Fujifilm X-series shooters, just as it is for those on other platforms. Although it isn’t always going to be your best bet, it performs better than the alternatives frequently enough that if you’ve the time required for its processing, it’s worth making your default option for high-ISO shots.

It’s also worth noting that even when it struggles with things like text, you usually have to be looking at the image very closely to spot the defects. Viewed from a more typical distance, a DeepPRIME XD shot is going to create the impression of more detail and crispness, even if it’s to some extent an artificial invention. When not pixel peeping, that can make the overall image feel better even when a more rigorous examination might tell a different story.

You have to bear your subjects, the viewer and how they’ll be looking at your images in mind when deciding which algorithm to use.

The point here is that you have to bear your subjects, the viewer and how they’ll be looking at your images in mind when deciding which algorithm to use. And we’ll restate that we’ve only shown results at default settings; you can always dial back the strength of an algorithm to improve the results for any given shot.

Overall, I’ve found DeepPRIME XD to be a very useful tool and I’m thrilled that it’s now available for X-Trans shooters too! If you’re not already a PhotoLab customer, I highly recommend checking out the free 30-day trial with your own images to see if it’s worth adding to your own digital darkroom.