Artificial intelligence enhances the performance of cameras and software. Noise-reduction algorithms reveal the original colors and unexpected details. Fake news or reality? Here’s how to check.
Noise increases at higher ISO settings. This series of shots of the same subject, made with a Nikon D850 at ISO settings ranging from 100 to 51,200, shows the gradual loss of color and detail starting at ISO 6,400, without any specific noise reduction processing.
Adobe has been offering AI-powered noise reduction in its image-processing software since 2023. DXO had already introduced DeepPrime in 2020, and Topaz Denoise in 2019. The 2026 versions of these AI-powered software programs offer improvements in both image quality and processing speed.
Nevertheless, one question always comes to mind whenever AI is used: fact or fiction? Do the colors and details produced by noise reduction match reality or not?
ISO, Noise, and AI
Noise increases at high ISO settings. Depending on the sensor generation and its size, noise may become noticeable as early as ISO 800 or only at ISO 5000. Regardless of the sensor, when we shoot at high ISO settings, it’s essentially for two reasons.
The first is aesthetic: we’re looking for a grainy look. The second is a workaround for shooting in low light while maintaining a fast shutter speed.
The shot made at ISO 100 (on the left) serves as a reference for the subsequent processing of photos taken with high ISO settings, such as this one captured at ISO 102,400.
Satisfaction with noise reduction depends on one’s memory of the scene and the subject.
This is because it’s rare to take a reference photo of the scene at a low ISO setting in order to compare image quality across the sensor’s sensitivity range. At very high ISO settings, image details become blurred, colors appear desaturated, and there is color shifting.
The recovery of details and colors by AI is so surprising that one might wonder if it’s not deceiving us. In an era of debates about the use of AI on social media, where misinformation competes with reality, how can we verify it? Here is a simple and effective procedure.
Organizing the shooting sequence
At low ISO settings, the rendering of details and colors most closely matches the subject being photographed. So let’s try the following exercise: take a series of RAW shots of the same subject while gradually increasing the ISO setting. The aperture value must be the same for all photos to ensure consistent optical quality across all shots.
This also means working in low-light conditions to avoid being limited by an unavailable shutter speed for very high ISO settings. For example, if the camera’s shutter speed is limited to 1/8000 s and the aperture is set to f/8, an ISO of 102,400 implies an aperture-shutter speed combination of 1/8000 s at ISO 102,400 for f/8, which is equivalent to ⅛ s at f/8 for ISO 100.
A tripod is essential to prevent camera shake, as shooting at a low ISO setting will likely require a slow shutter speed.
Comparison of processed images
The images are then processed in Lightroom, DxO PhotoLab, PureRaw, or alternatives such as Topaz DeNoise. Note, however, that processing time depends on the performance of the computer’s graphics card.
We recommend an M-series processor from Apple and an NVIDIA RTX graphics card on a PC.
The low-ISO image, which has not been denoised, serves as a reference for comparison with the denoised photos taken at high ISO. The amount of denoising can be adjusted as needed.
In its early versions of AI-powered noise reduction, Lightroom generated a noise-reduced DNG file. Today, noise reduction is built into the native RAW file. Processing time depends on the power of your graphics card.
The time indicated is approximate.
DXO offers its DeepPrime noise reduction feature in its PhotoLab and PureRaw software. The luminance slider is set to 40 by default. The preview window does not accurately reflect the final result. Only by exporting to DNG can you verify that the noise reduction level is set correctly.
Before and After in Lightroom
On the left is a cropped section of a Nikon D850 RAW file shot at ISO 102,400 without any processing. In the middle, AI-powered noise reduction with the slider set to 50 removes most of the noise and restores detail and color well, particularly in the reds. On the right, with the slider set to 80, the noise is even better controlled.
Before and After with DXO
On the left is the same selection from the RAW file. On the right is the noise reduction applied by DXO DeepPRIME XD3, DXO’s latest processing technology. With the slider set to the default value of 40, noise is less noticeable than with Adobe’s AI, detail recovery is more satisfactory, and colors are more vivid.
DXO vs Adobe
With the default settings (sliders set to 40 for DXO DeepPrime and 50 for Lightroom), DXO DeepPrime’s noise reduction is more effective than Adobe’s, particularly in terms of color reproduction.
The red here is more faithful to the original scene.
From 100 ISO to 102,400 ISO
Let’s compare the shot taken at ISO 100 (center) with the one taken at ISO 102,400, processed in Lightroom (left) and in DxO PhotoLab (right). The DxO processing is more accurate in terms of color rendering. Noise is better controlled. However, we can see that the fabric details have been lost at ISO 102,400 in both versions.
Verdict
How much noise reduction you prefer is a matter of personal taste. We believe that a hint of noise adds a bit of texture to the image. A photo that’s been smoothed out too much looks artificial.
It’s up to each person to find the right balance between effective noise reduction and acceptable smoothing of details.
The verdict? DxO outperforms Adobe in color reproduction and noise reduction. A comparison with other software would be interesting, but is beyond the scope of this article. It’s up to you to try it out. In any case, AI impresses us with its progress in realism, even if there is still room for improvement, particularly in detail reproduction.
We don’t quite reach the quality of an image taken at the sensor’s native ISO, but we come close. We encourage you to test this out. AI’s progress toward realism is certainly impressive, approaching, though not yet fully matching, the quality of images captured at the sensor’s native ISO. While the level of detail reproduction still offers room for improvement, the results are remarkably close.
Discover the professional photography courses at Spéos
Spéos photography school offers professional photography programs in 1 year, in 2 years, short photography programs (in 2 months and in 5 months), as well as photography workshops.
The long courses to become professional photographers allow you not only to master all the photographic techniques and its vocabulary (blurs, hyperfocus, sharpness zone, depth of field, backlighting, focal length, shutter release, autofocus, wide-angle, rule of thirds, etc.), but also all the stages of shooting and image processing.
Visiting the school allows you to discover the premises, the studios and the equipment, and is undoubtedly the best way to familiarize yourself with your future way of working. This is why, in addition to the open days, Spéos offers throughout the year personalized visits by appointment to come and discover the school with a member of the team.
Text and photos: Philippe Bachelier, teacher of Printing techniques at Spéos