Topaz DeNoise AI
Version 3.7.2 · Topaz Labs · Windows 10 / 11 (64-bit) · ranked #146 in the library
A dedicated noise reduction tool that uses trained models to tell grain apart from genuine detail, cleaning up high ISO and low light photographs while keeping the fine texture a blanket denoise would smear away.
Read the installation section below before you start. The order of the steps matters on this release.
About this release
Overview
DeNoise does one thing that general editors do poorly. Noise reduction has always been a trade against detail, because the traditional approach cannot tell the difference between random sensor noise and the fine texture of the subject, so turning one down takes the other with it. The trained model here was built specifically to separate the two, which is how it clears the grain out of a high ISO or low light frame while keeping the eyelashes, the fabric weave and the feather detail sharp.
It offers several models for different situations, because a clean image shot at high ISO, a low light frame full of luminance noise, and a heavily compressed file with colour blotching are different problems. Each model previews on the actual image before you commit, so you are choosing based on your frame rather than a general recommendation, and a detail recovery control lets you push texture back after the noise is gone.
It runs standalone and as a plugin into a host editor or catalogue application, so it drops into an existing photography workflow as one step rather than replacing it. It reads raw files directly, which is the right place to denoise because the data is richest before other processing, processes on the GPU, and handles a batch of frames from the same shoot in one pass. The models are bundled so it works offline.
What it does
Feature breakdown
Detail preserving noise reduction
A trained model that separates sensor noise from genuine texture, so fine detail survives the cleanup.
Multiple models per situation
Different models for clean high ISO, heavy low light noise and compression artefacts, previewed on the real image.
Raw file input
Reads raw files directly, denoising at the richest point in the pipeline before other processing.
Standalone and plugin operation
Runs on its own or as a plugin into a host editor or catalogue application.
Detail recovery control
Push texture back after the noise is removed, for control over the balance on a given frame.
Batch processing on the GPU
Process a whole shoot in one pass, with the work running on the graphics card.
Changed in this build
Version 3.7.2
- New model tuned for extreme low light frames.
- Faster processing on recent GPUs with better memory handling.
- Improved colour noise handling on compressed sources.
- Preview comparison rebuilt for quicker model selection.
- Additional raw profiles for recent camera bodies.
Installation
Follow these in order
- Go offline.
- Extract the archive.
- Disable real time protection.
- Run the installer and complete setup.
- Copy the included model files into the models directory if setup did not place them.
- Launch the application, confirm the models are listed, then re-enable protection.
Release notes
From whoever packed it
- The models are bundled here. Without them the application starts but has nothing to run.
- Block outbound access. An online check reverts the build to watermarked trial output.
- For plugin use, point the installer at your host editor's plugin folder during setup.
System requirements
Minimum and recommended
| Operating system | Windows 10 version 22H2 or Windows 11, 64-bit |
|---|---|
| Processor | Modern quad core or better |
| Memory | 8 GB minimum, 16 GB recommended |
| Graphics | GPU with 4 GB VRAM minimum, 6 GB or more preferred |
| Storage | 3 GB for the application and models |
Questions about this release
Answered before you ask
- Does it keep detail?
- That is the whole point. The model separates noise from texture, so fine detail survives the cleanup.
- Can I use it inside my photo editor?
- Yes, the plugin build hands results back to a host editor. It also runs standalone.
- Is the output watermarked?
- No, provided the application stays offline and the bundled models are in place.
Download
Pick a mirror