Noise reduction that separates grain from detail rather than blurring both, with several model paths for different kinds of noise and a raw pass that runs before demosaicing.
Conventional noise reduction is a smoothing operation, and smoothing cannot tell the difference between sensor noise and the fine detail of feathers, fabric or foliage. Push it far enough to clean a high sensitivity frame and the subject turns waxy. DeNoise AI works differently: the models were trained on matched pairs of noisy and clean frames, so the engine has learned what detail looks like under noise and reconstructs it rather than averaging it away.
There are several model paths and they suit genuinely different problems. The standard model handles ordinary high sensitivity noise. The clear model is more aggressive and suits heavy noise where some detail loss is acceptable. The low light model handles the colour blotching that appears in very dark frames. The severe noise model deals with the extreme cases from very small sensors and very long pushes. Each has strength and detail recovery sliders, and the comparison view shows all of them against the original at once so you can pick by looking rather than guessing.
The raw path is the technically interesting one. Running noise reduction directly on raw data before demosaicing produces a cleaner result than doing it afterwards, because the demosaic step spreads noise across neighbouring pixels and makes it harder to identify. Feeding it raw gives the engine the sensor data as it was recorded.
It runs standalone and as a plugin inside the common host editors, and it batches, which is where the time actually goes on an event or wildlife shoot of several hundred frames. Output preserves the full bit depth and the colour profile, masking allows the correction to be applied selectively, and processing runs on the graphics card with a processor fallback for machines without a supported one.
Read the whole list before you start. Most of the reports that come into the desk about this title turn out to be a step taken out of order.
| Processor | Intel or AMD, 64 bit, with AVX2 support |
|---|---|
| Memory | 8 GB minimum, 16 GB recommended |
| Graphics | GPU with 4 GB VRAM for the accelerated path |
| Storage | 6 GB free |
| Display | 1920 by 1080 or higher |
Judge the result at full size, never in the fitted preview. Noise reduction always looks fine when it is scaled down.
Feed it raw where you have it. The improvement over processing an already developed file is larger than any slider setting.
| Title | Topaz DeNoise AI |
|---|---|
| Developer | Topaz Labs |
| Version | 3.7.2 |
| Category | Photo & Graphics |
| Release type | Full version |
| Operating system | Windows 11, Windows 10 (64 bit) |
| Architecture | x64 |
| Interface language | English |
| Archive size | 1.42 GB |
| File name | topaz_denoise_ai_3.7.2_x64_full.rar |
| Shelf updated | 5 days ago |
| Total downloads | 163,304 |
The full desktop build of Photoshop with the generative model stack left switched on, including Flux Kontext Pro and Gemini Nano Banana routing.
The vector drawing standard, with the full type engine, live shape tools and the generative vector pass that produces editable paths rather than flat art.
Catalogue driven photo management and raw development for people who shoot in volume and need the whole library in one place on their own disk.
Vector illustration, page layout, photo editing and font management in one suite, with the tracing engine and print production tools that sign and apparel shops run on.