Adobe Photoshop 2026
The complete 2026 raster editor with the full generative model set switched on.
Noise reduction that separates grain from detail using trained models rather than smoothing, recovering usable images from high sensitivity frames that conventional tools turn to mush.
Overview
Conventional noise reduction works by smoothing, and smoothing cannot tell the difference between sensor noise and fine detail, so it removes both. The result is the familiar waxy look where the noise is gone and so is every eyelash and fabric weave in the frame. This program takes a different route, using models trained on matched pairs of noisy and clean images to work out which parts of the frame are noise and which are detail.
In practice that means a frame shot at a sensitivity you would normally throw away comes back with the grain removed and the texture still present. It is the difference between an unusable file and a printable one, and it is why the program found its way into the workflow of anyone who shoots in poor light.
Four models are included and the automatic mode picks between them. The standard model handles most cases. The low light model deals with the mixed noise and colour blotching that appears on very long exposures and very high sensitivity. The severe noise model goes further for the worst cases at some cost in detail. The raw model works on undemosaiced data, where noise reduction is most effective because it happens before the colour interpolation stage smears the noise around.
Everything is previewed at full resolution with a split view and a comparison grid showing all four models against the original, so the choice is made by looking rather than by guessing. Sharpening and detail recovery sliders sit alongside so the recovered detail can be brought back up without amplifying what was removed.
It runs standalone and as a plugin inside the major photo editors, with batch processing for pushing a whole shoot through the same model. This build is licensed with the trial watermark removed.
Feature set
Trained to separate sensor noise from real texture rather than smoothing both together.
Standard, low light, severe noise, and a raw model that works before demosaicing where noise reduction is most effective.
All four models previewed against the original at full resolution so the choice is made visually.
Sharpening and recovery controls that bring texture back without amplifying what was removed.
Push an entire shoot through one model with consistent settings and no per file attention.
Works on its own or inside the major photo editors, fitting into an existing workflow.
Changes in this build
Installation
Hardware
| Processor | Intel or AMD, 64-bit, with AVX2 support |
|---|---|
| Memory | 8 GB minimum, 16 GB recommended |
| Graphics | GPU with 4 GB VRAM recommended, processor fallback available but much slower |
| Storage | 3 GB free space |
| Display | 1920 x 1080 or higher for the comparison view |
| System | Windows 10 version 21H2 or Windows 11 |
Package
| Title | Topaz DeNoise AI |
|---|---|
| Version | 3.7.2 |
| Developer | Topaz Labs |
| Category | Photo Editing |
| Platform | Windows 10 / 11 (64-bit) |
| Package size | 1.34 GB |
| Interface languages | English |
| Release identifier | BS-37200001 |
| Listed | 3 years ago |
| Last rebuilt | 6 days ago |
| Total downloads | 163,304 |
Questions
Start with automatic, then check the comparison grid. Most cameras settle on one model for most of your work.
Yes, and the raw model is the strongest option because it runs before demosaicing, where the noise has not yet been spread across channels.
It falls back to the processor, correctly but far more slowly. A batch will take hours rather than minutes.
The recovery sliders default to a level suited to heavier noise. Reduce them once the noise itself is handled.
The complete 2026 raster editor with the full generative model set switched on.
The desktop catalogue application photographers actually use for volume work.
A dedicated image upscaler that enlarges photographs several times over while rebuilding the detail a plain resize would smear.
The last standalone perpetual release of the photo catalogue and develop application.