Overview
Topaz previously sold noise reduction, sharpening and enlargement as three separate programs, and the practical problem with that arrangement was ordering. Sharpen before denoise and you amplify the noise. Enlarge before denoise and you enlarge the noise as well. Photo AI collapses all of them into one program that examines the image, works out which problems it actually has, and applies the right corrections in the right sequence without you having to know the theory.
This build has the complete model set installed rather than downloaded, and the licensing is settled ahead of time so there is no watermark and no restriction on output. Everything runs on the local machine with no image sent anywhere, and the program works as a standalone application, as a plug-in inside the major raster editors, and as an external editor from the catalogue programs.
The automatic pass
Open an image and the program analyses it, detects faces, estimates the noise level and the amount of blur present, and enables the modules it thinks are needed with starting strengths already set. On a well exposed photograph it often does nothing at all, which is the correct answer and a good sign that the detection is honest rather than applying processing for its own sake.
Everything it decided can be overridden. Each module can be switched on or off, its model changed and its strength adjusted, and the changes preview immediately in the split view. The autopilot preferences let you bias those automatic decisions permanently, which is worth setting once if you find the defaults consistently a little strong or a little weak for the material you shoot.
What each module does
Denoise handles sensor noise at high sensitivity settings, with models for normal, strong and extremely noisy material plus a raw specific path that works on the sensor data before demosaicing, which produces a noticeably better result than working on a processed file. Sharpen addresses blur from motion, from missed focus and from lens softness, with a separate model for each because they are different problems and correcting one as if it were another produces artefacts.
Face recovery rebuilds facial detail in portraits where the subject is small in the frame or where the original was low resolution, and it has a strength control because it can be overdone. Upscaling handles enlargement using the same approach as the dedicated enlargement program, and lighting and colour balance modules cover the broader corrective work.
Workflow and batch
As a plug-in it returns the result as a new layer inside the host editor, so the original stays intact and the correction can be masked back where it is too strong. From a catalogue program it operates as an external editor with the round trip handled automatically. Standalone, it accepts a folder and processes everything with either the automatic decisions per image or a single fixed setting applied to all of them.
Raw files are supported directly, which matters because the raw denoise path is meaningfully better than processing after conversion. Output formats cover the usual raster set plus the layered format for handing off into a raster editor.
What you get
- Automatic analysis that selects the needed corrections and their order
- Denoise with normal, strong and extreme models plus a raw specific path
- Separate sharpening models for motion blur, missed focus and lens softness
- Face recovery with adjustable strength for small or low resolution subjects
- Upscaling using the dedicated enlargement models
- Lighting and colour balance correction modules
- Split view preview with per module override of every automatic decision
- Batch processing with per image automatic settings or one fixed profile
- Standalone, plug-in and external editor operation, all fully local
Inside the archive
- Topaz Photo AI installer
- Complete model set, preinstalled rather than downloaded
- Plug-in components for the supported host editors
- Readme covering the autopilot preferences and the raw workflow
System requirements
| Operating system | Windows 10 version 22H2 or Windows 11, 64-bit |
|---|---|
| Processor | Intel or AMD x64 with AVX2 |
| Memory | 16 GB recommended |
| Graphics | 6 GB VRAM strongly recommended, processor fallback available but slow |
| Storage | 10 GB free including the model set |
| Display | 1920 by 1080, colour managed panel recommended |
Installing it
- Unpack the archive to a local folder and suspend real time protection.
- Run setup as administrator and allow the model set to install with the program.
- Launch once and set the processing device in preferences, choosing the graphics card if one is available.
- Open the autopilot preferences and set your bias for face recovery and sharpening strength before processing a batch.
- Run one image through to confirm the device selection works before queueing a folder.
Mirrors
| Route | Region | Note | State |
|---|---|---|---|
| Direct, primary | Europe | No wait, resumable | Online |
| Direct, secondary | North America | No wait, resumable | Online |
| Torrent magnet | Global | Optional at this size | Online |
Release history
- Raw denoise path extended to further camera formats
- Face recovery strength now defaults lower on group photographs
- Fixed batch processing stalling on files with embedded colour profiles it did not recognise
- Sharpening model selection improved on images with mixed blur types
- Plug-in components updated for the current host editor releases
- Autopilot preferences added for permanent bias adjustment
- Model set bundled into the archive instead of downloaded on first launch
Questions about this release
Why does it sometimes apply nothing?
Because the image did not need anything. Honest detection is the point, and a program that always applies processing is not doing you a favour.
Should I feed it raw files or processed ones?
Raw, where possible. The denoise path works on the sensor data before demosaicing and the result is meaningfully better.
Does it replace the dedicated enlargement program?
For most work yes, since it uses the same models. The dedicated one gives finer control over enlargement specifically.
Is anything sent to a server?
No. All processing runs locally and the model set is installed with the program.
Comments
Like that it does nothing when the photo is fine. Other tools always find something to smear.
Runs as a plug-in and comes back as a layer so I can mask it. That is the workflow I wanted.
Comments are read before they appear. If a build stops working, say so here and it gets rebuilt rather than quietly left up.
Raw denoise on high sensitivity concert shots is the best I have seen. Kept detail I assumed was gone.