Reconstruction sold as noise reduction

What separates it from a slider

Classical noise reduction analyses the local neighbourhood and smooths what looks random while trying to protect what looks like an edge. It has no idea what the picture contains, so it faces the unavoidable trade from the earlier lesson: luminance noise and fine texture are the same thing at the pixel level.

A learned denoiser was trained on pairs of noisy and clean images of real scenes. It has learned what clean photographs of fabric, foliage, skin and brick look like, so it does not smooth: it predicts the clean version. That is why the results are so much better, and why they carry a different risk.

The headline improvement is genuine: on a high-ISO raw file, the good implementations give the equivalent of two to three stops. An ISO 12800 photograph comes back looking like ISO 1600.

Why several of them are raw-only

Adobe’s Denoise, DxO’s DeepPRIME and several others work on the raw file before demosaicing.

The reason is that noise in raw data has known statistics: it is dominated by photon shot noise, which is Poisson-distributed and therefore predictable from the signal level, plus a read noise component characteristic of the sensor. A model trained on that distribution has a well-defined problem.

Once the file has been demosaiced, white-balanced, tone-curved and JPEG-compressed, the noise has been through several non-linear transformations and is correlated between neighbouring pixels. It no longer matches what the model was trained on. That is why denoising a JPEG is markedly less effective than denoising the raw, and why these tools often refuse JPEGs entirely rather than doing a poor job.

It is also a strong practical argument for shooting raw in low light, separate from the exposure argument.

The characteristic failures

Fine repeated texture is replaced. Fabric weave, distant foliage, hair, a wire fence, small text on a sign. The model produces texture that is statistically right and locally wrong. On a portrait nobody cares; on a photograph of a document it matters entirely.

Waxy skin at high strength. The model’s prior for skin is smooth, so pushed hard it returns smooth skin. This is the same plastic look as over-aggressive classical denoising, arrived at differently.

Small faces become different faces. Where a face occupies few pixels, the model fills in from its learned prior of faces. Group photographs are where this shows.

Stars and specular points vanish. They look like impulse noise. Astrophotographers turn these tools off for that reason.

Detail that is not in the training distribution is handled badly: unusual materials, scientific imagery, microscopy, infrared.

Deblurring, which is weaker than it appears

The unrecoverable-losses lesson explained why blind deconvolution is ill-posed. Learned deblurring does not solve that; it sidesteps it by generating a sharp-looking image consistent with the blurred input.

On mild, simple blur it produces useful results. On anything substantial it produces a confident, sharp, plausible image whose fine detail is invented. The giveaway is usually over-clean edges with a slight plasticity, and text that reads as text from a distance and not close up.

Treat any “unblur” or “enhance” feature as generation, and apply the same test as everywhere else: does the content of this region matter as a record?

Free tools

• darktable’s profiled denoise uses measured noise profiles for specific camera and ISO combinations, so it knows the statistics of your sensor’s noise at that setting. It is not a learned model and on many files it is competitive with the paid learned ones, particularly its wavelet and non-local means modes used together: chroma with non-local means, luminance with wavelets.
• RawTherapee’s noise reduction is also strong and highly configurable.
• Neat Image and DxO are the paid standards.
• G’MIC inside GIMP has a large set of denoisers including some learned ones.
• Several Real-ESRGAN variants denoise as a side effect of upscaling at 1×, which is a usable free trick: upscale 2× with denoising and downscale back.

darktable’s approach deserves particular mention because it is the honest one: it uses a measured profile of your actual sensor rather than a prior about what photographs look like, so it removes noise without inventing content. It is less dramatic and more truthful.

How to use a learned denoiser well

1. Apply it to the raw, early, before anything else.
2. Use less than the default. Most implementations default high. Half strength usually keeps the texture and removes the objectionable noise.
3. Compare at 100 per cent against the undenoised version, looking specifically at fabric, hair and foliage.
4. Mask it off areas that must stay authentic, where the tool allows. Some do not, in which case process twice and blend layers by hand.
5. Add grain back if the result looks synthetically clean next to the rest of a set.
6. Sharpen after, never before.

The statement to hold

A learned denoiser is a reconstruction tool sold as a cleaning tool. For a wedding, a portrait, a concert, a family photograph (almost all paid photographic work) it is a large improvement and the reconstruction is invisible and harmless.

For a photograph that will be used as evidence of what something looked like (a document, an identifying detail, a scientific measurement, a news image, an insurance claim) the output is not a cleaned version of the original. It is a new image consistent with the original, and it should be labelled as processed or not used at all.

That distinction is not a slider setting. It is a decision about what the picture is for, and it is made before you press the button.