AI Noise Reduction for Photogrammetry in 2026: Lightroom Denoise Tested

Back in 2023 I wrote about my RAW processing workflow for photogrammetry, compared AI and traditional noise reduction, and then tested what noise reduction does to a 3D reconstruction. The conclusion then was simple: the best results came from RAW files converted without any extra noise reduction. Noisy, but with all of the detail intact and the soundest 3D structure.

AI denoising has come a long way since, and it now sits inside Lightroom as a single button. So it seemed like a good time to check whether that conclusion still holds, this time with real check measurements instead of a visual judgment.

The Short Version

• I ran Lightroom's AI Denoise on the three noisiest sets from my low light roof test and compared them against the same sets without it.
• Denoise made the photos look clean, and made precise measuring points noticeably easier to place.
• It did not improve the reconstruction. Every denoised set produced fewer dense points, and the check measurements were no more accurate.
• It took about two minutes per set of 35 photos, which scales to roughly an hour for a 1,000 photo job.
• The 2023 verdict stands: for measurement, skip it. For a clean looking deliverable, it is fine.

The 2023 Workflow in 2026

The basics of the RAW workflow have not changed. Develop in Lightroom, push highlights down and shadows up to recover detail, set an accurate white balance, keep adjustments modest, and sync identical settings across the whole set. Export at full resolution, keep all metadata so the GPS positions survive, and never add output sharpening or watermarks.

Two things are different. The first is why you would shoot RAW at all. In 2023 the answer was mostly high contrast scenes at sunrise and sunset. After the low light series this autumn, I would add any borderline light, because RAW gives you far more room to recover a dark set than a JPEG does.

The second is AI noise reduction being one click away, which makes it tempting to apply to everything. That is exactly what this test is about.

Why AI Denoise Is a Risk in Photogrammetry

Traditional noise reduction is subtractive. It smooths the image, taking away grain along with some fine detail. AI noise reduction works differently. It has learned what real detail is supposed to look like and rebuilds the image accordingly, which is why the results look so clean.

For a single photo that is a pure win. For photogrammetry it is a potential problem, and I pointed it out back in 2023. Every photo in a set is denoised separately, so the rebuilt detail can come out slightly different in each frame. Photogrammetry depends on finding the same feature in many overlapping photos. Detail that varies between frames is not detail to the software, it is inconsistency.

The Test

The photos are from Part 2 of the low light series: a Mavic 4 Pro orbit of a residential roof in heavy overcast, all at 1.7 m/s, about 35 photos per set. I used the three sets where noise was the main problem rather than blur: ISO 6400, 3200 and 1600. Below that, the photos were limited by motion blur, and no noise reduction can fix blur.

Each set was processed twice in Lightroom with identical settings, once as before and once with AI Denoise added. I used a strength of 70, which to my eye removes virtually all of the bad looking noise. Same JPEG export, then the same default medium settings in Pixpro.

To judge the results I used the method from the measurement validation article: two objects measured with a tape on site, a 3,000 mm zebra crossing stripe at the edge of the scene and a 1,210 mm roof window side near the center, both measured in Pixpro with projections.

The Results

What the Numbers Show

Fewer points, every time. Every denoised set produced a smaller dense cloud. The biggest drop was at ISO 6400, about 6 percent, which is the set where Denoise had the most work to do. That fits how noise reduction works: smoothing away grain also smooths away some of the fine texture that dense matching relies on.

No gain in accuracy. Of the six check measurements, four got clearly worse with Denoise, one stayed within a millimeter, and one got better, the zebra at ISO 6400, while the window in the same set went the other way. At ISO 1600, the cleanest set, both measurements moved further from the truth. These are single runs and the differences are a few millimeters to a few centimeters, so I would not read them as proof that Denoise damages a model. But there is clearly no improvement.

The scale error did not care. ISO 3200 was off by about +2 percent on both objects before, and it still is after. As explained in the validation article, matching errors like that come from scale, most likely the drone's GPS, and processing the photos differently does nothing about it.

The cloud looked about the same. The denoised dense clouds looked good, but not markedly better than the noisy ones.

Easier to Measure, Not More Accurate

The one clear benefit was in measuring. On the clean, denoised photos, placing projections was noticeably easier. Corners were crisp, edges were obvious, and there was no grain to guess through.

And yet the measurements did not get better. My best explanation is that a crisp corner on a denoised photo is partly the denoiser's best guess at where that corner should be. It looks precise, and it is easy to click on, but looking precise and being in exactly the right place are not the same thing. It is a small trap: the photo gives you confidence that the numbers do not back up.

The Time Cost

Denoise took about two minutes per set on a decent PC, so about three and a half seconds per photo. For a 35 photo roof that is nothing. For a 1,000 photo site it is close to an hour of extra processing before the reconstruction even starts, for no measurable gain.

Verdict

The 2023 conclusion holds. For measurement work, skip AI noise reduction and process the RAW files without it. The noisy set reconstructs at least as well and measures at least as accurately, and it costs you nothing in time.

Where Denoise does make sense is presentation. If the client will look at the textured mesh rather than measure it, a clean texture is worth a couple of minutes. In that case the 2023 advice still applies: measure on the version processed without noise reduction, and show the client the clean one.

Conclusion

AI noise reduction is impressive, and on a single photo it is close to magic. Photogrammetry is not a single photo, though. It is dozens or hundreds of photos that have to agree with each other, and a tool that rebuilds each one separately does not help with that, even when every frame looks perfect.

Three years and a much better denoiser later, the answer is the same. Clean looking photos are nice to work with. They are not more accurate.

Frequently Asked Questions