Local Windows utility
AI Enhance
A photo that is too small for what you want to do with it, made twice the size by a model that reconstructs detail rather than stretching pixels. It runs on your own CPU, with no account, no credits and no per-image charge — and it is not fast, which PhotoTray tells you before it starts rather than after.
Upscaled copies land in the AI Enhance folder — your originals are never modified.
What it's for
Real jobs — AI Enhance
Not feature bullets — the actual situations people open this for.
- Rescuing a small profile photo or an old scan that looks soft at full size.
- Making a product shot big enough to print without it going to mush.
- Enlarging a photo someone sent you through a messaging app that shrank it.
- Getting an old low-resolution image up to a size a modern screen does not embarrass.
What it does
Everything AI Enhance handles
Everything below runs on your own machine, on as many files as you give it.
Twice the width, twice the height
Real-ESRGAN reconstructs detail rather than interpolating between the pixels it already has, so edges stay edges instead of turning into staircases.
The model is in the installer
Nothing downloads on first use and nothing needs a connection. The 64 MB network ships with the app and is checksum-verified before it is ever executed.
No account and no credits
Every comparable tool online meters this by the image. Here it runs on your CPU, as many times as you like, for nothing.
An honest time estimate
CPU upscaling is slow. PhotoTray measures the photos you actually added and shows what they will take before you queue them, rather than leaving you watching a bar.
Batch, like everything else
Point it at a folder and every photo becomes its own job — with the caveat that a folder of full-size photos is an overnight run, not a coffee-break one.
It can be a rule
Upscaling is automatable, so a Watch Folders rule can chain it after a conversion — the SD-card preset does exactly that.
How it works
From drop to done
- 1
Drop in a photo
Or a folder, if you have the time to spare. JPG, PNG, WebP, GIF, TIFF and AVIF.
- 2
Read the estimate
PhotoTray works out what your photos will cost from their pixel counts and your CPU's core count, and shows it before anything is queued.
- 3
Let it run
Each photo becomes its own job in the queue. There is no connection involved at any point.
Under the hood
The engines, named
The model, named — and where it runs.
- Upscaler
- Real-ESRGAN x2 (ONNX weights) on onnxruntime — CPU
- Scale
- 2× — the factor the shipped network was trained for
- Input
- JPG, PNG, WebP, GIF, TIFF, AVIF
- Model delivery
- bundled in the installer, SHA-256 verified before use
- Platform
- Windows 10/11 (x64) — CPU only, no GPU required
CPU-only is a deliberate trade, not an oversight: it means no GPU requirement, no driver problems and no machine where the feature simply refuses. The cost is speed — measured at 2× on eight cores, roughly 36 seconds per input megapixel, so a 12-megapixel phone photo is around seven minutes. There is no 4×: Real-ESRGAN upsamples by exactly the factor it was trained for, so 4× would be a second network and about four times that cost. Running a 2× pass twice is not the same thing and is not offered.
The rest of the belt
It ships with five more
Questions
AI Enhance, answered
Get the whole tray — AI Enhance
AI Enhance comes with PhotoTray — free for Windows 10 and 11, with the other five alongside it.