Overview
Traditional sharpening increases local contrast at edges. It makes an image look crisper from a distance and it makes the halos and the noise worse when you look closely, and it cannot recover anything that was not captured. Sharpen AI works differently: the models were trained on pairs of soft and sharp versions of the same scene, so instead of exaggerating what is there, the program reconstructs what a sharp version would plausibly have looked like.
The important design decision is that softness has three distinct causes and each needs a different answer. A photo that missed focus, a photo blurred by camera or subject movement, and a photo that is simply not as crisp as the lens should have given are three different problems, and the program has a separate model for each. Choosing the right one is most of getting a good result. This build is posted with the licensing handled and the model set included offline rather than downloaded on first run.
The three models
Focus is for a shot where the focus point landed behind or in front of the subject. It is the most aggressive of the three and the one most likely to invent detail, so it wants a careful eye at full magnification. Within about a stop of correct focus the results are frequently good enough to rescue a shot that would otherwise be deleted.
Motion handles blur from movement, whether the camera moved or the subject did. It works out the direction and extent of the smear and reverses it, which is a genuinely difficult problem and the model handles moderate cases surprisingly well. Long exposures with large movement are still beyond it.
Standard, sometimes called too soft, is for images that are technically in focus but lack bite, which covers diffraction at small apertures, a soft corner on a cheap lens, and the general mush of an image that has been resized. It is the gentlest model and the safest default when you are unsure.
Control and masking
Every model has a strength slider and a separate suppression control for the noise that sharpening amplifies. The preview is split, comparing before and after at full resolution rather than on a downscaled proxy, which matters because sharpening artefacts are invisible until you look at actual pixels.
Selective application is the feature that separates a usable result from an obvious one. The masking brush limits the effect to the parts of the frame that need it, typically the eyes and the face in a portrait, leaving the background and the smooth areas untouched. Applying a strong sharpen across a whole frame is the fastest way to make an image look processed.
Batch work and how it fits a workflow
Batch processing runs a folder through a chosen model and setting, which is the right approach for a set shot under the same conditions. Autopilot inspects each image and picks a model and strength on its own, and it is right often enough to be useful as a starting point that you then override on the few that need it.
It works as a standalone application and as a plug-in from Photoshop, Lightroom and the other common hosts, taking a layer or a raw develop and returning the result. Sharpening should always be the last step before output, after any resizing and grading, because doing it early means every subsequent operation works on top of the artefacts.
What to expect and what not to
It cannot recover an image that missed by a mile. If the subject is unrecognisable, no model will produce a real face, though it may produce a convincing one, which is a different and more dangerous thing. On genuinely borderline images it is remarkable, and on the many images that are just slightly soft it is the fastest quality improvement available.
Processing is GPU accelerated and a card with 6 GB or more turns a several second wait into an immediate one. On CPU only it still works, at maybe fifteen times slower, which makes batch work impractical but single images fine.
What you get
- Separate trained models for focus, motion and general softness
- Strength and noise suppression controls per model
- Full resolution split preview rather than a downscaled proxy
- Masking brush for selective application
- Autopilot model and strength selection
- Batch processing across folders
- Standalone application and host plug-in
- Raw file support alongside JPEG, TIFF and PNG
- GPU acceleration with CPU fallback
- Models bundled offline rather than downloaded on first launch
Inside the archive
- Sharpen AI standalone application, offline installer
- Complete model set, preinstalled
- Plug-in components for the common hosts
- Readme covering the plug-in registration step
System requirements
| Operating system | Windows 10 version 21H2 or Windows 11, 64-bit |
|---|---|
| Processor | Quad core with AVX support |
| Memory | 8 GB minimum, 16 GB recommended |
| Graphics | 4 GB VRAM minimum, 6 GB or more for responsive processing |
| Storage | 4 GB free |
| Display | 1920 by 1080, colour accuracy helps when judging results |
Installing it
- Unpack the archive to a local folder.
- Turn off real time protection for the install.
- Run setup as administrator.
- Register the plug-in for your host application from the setup options, not afterwards.
- Open the program once and confirm the model list is populated without it trying to download anything.
Mirrors
| Route | Region | Note | State |
|---|---|---|---|
| Direct, primary | Europe | No wait, resumable | Online |
| Direct, secondary | North America | No wait, resumable | Online |
| Torrent magnet | Global | Optional | Online |
Release history
- Motion model retrained, better on subject movement rather than camera shake
- Faster GPU processing on current cards
- Fixed masking brush edge behaviour at high zoom
- Autopilot selection accuracy improved
- Raw support extended to more camera bodies
- Major release with rebuilt models across all three types
- Batch processing reworked
Questions about this release
Which model should I use?
Standard if the image is in focus but soft, Focus if the focus point missed, Motion if something moved. Autopilot guesses reasonably and you can override it.
Does it need an internet connection?
No. The models are bundled in this build rather than fetched on first launch.
Can it fix a badly out of focus shot?
No. It works within roughly a stop of correct focus. Beyond that it will invent detail that was never there.
Where in my workflow should it go?
Last, after resizing and grading. Sharpening early means everything after it operates on the artefacts.
Comments
Mask the eyes only and it looks natural. Whole frame at full strength looks processed, as always.
Models installed offline, no download on launch. Runs in about two seconds on a 4070.
Comments are read before they appear. If a build stops working, say so here and it gets rebuilt rather than quietly left up.
Motion model saved a set of flight shots at 1/200. Not perfect but usable, which they were not before.