Topaz DeNoise AI

About this release

Every conventional noise reduction tool faces the same problem: noise and fine detail look alike to an algorithm working on local statistics, so removing one damages the other. The result is the smeared, waxy look that anyone who has pushed a slider too far will recognise. This tool works from models trained on large numbers of clean and noisy image pairs, so it has learned what feathers, fabric, foliage and pores look like and can tell them apart from sensor grain.

The practical effect is that images shot at high sensitivity become usable rather than merely acceptable. A concert frame at a very high sensitivity setting, an indoor shot at dusk, wildlife at the long end of a lens in poor light: these come back with the grain gone and the texture still there, which changes what a photographer is willing to shoot in the first place.

Several models are provided because different noise behaves differently. A standard model for ordinary sensor noise, a clearer model that leans towards preserving fine texture, a low light model for the extreme end of the sensitivity range, and a severe noise model for the images that would otherwise be discarded. The comparison view puts them side by side on the actual image so the choice is made by looking rather than by guessing.

It reads raw files directly, which is where noise reduction belongs in the pipeline, and it also runs as a plug in from the common host editors for anyone who prefers to keep the rest of their workflow. Batch processing takes a folder through the same settings, and processing runs on the graphics card so a full resolution frame comes back in seconds rather than in the minute a processor only pass would take.

What it does

Trained noise models

Detail and grain are separated by a model that has seen both, rather than by local statistics.

Four model choices

Standard, detail leaning, low light and severe noise, chosen by looking at the actual image.

Side by side comparison

Model and setting choices are judged on the real frame at full resolution.

Direct raw handling

Reads raw files so noise reduction happens before development rather than after.

Graphics card processing

Full resolution frames return in seconds instead of minutes.

Batch and plug in modes

Run a folder through one setting, or work from inside an existing host editor.

Changed in this version

• Faster processing on current generation graphics cards.
• Improved results on very high sensitivity frames with colour blotching.
• Raw format support extended to recent camera bodies.
• Batch queue reports per file progress and estimated time.
• Fixed a halo artefact along high contrast edges at strong settings.

System requirements

Install order

1. Unpack the archive and run the installer.
2. Choose whether to install the plug in for your host editor alongside the standalone application.
3. Enable graphics card processing in preferences after checking the adapter is detected.
4. Test the models on one representative frame before committing a batch.

Worth knowing before you start

File details