sqwartHome

AI Photo Enhancer

Clean up compressed, noisy and soft photos with an AI model that runs on your device.

  • Free
  • No upload
  • No watermark
  • No sign-up
  • AI runs on your device
How it works

How to use the AI photo enhancer tool

  1. 1

    Choose a photo, ideally a small or compressed one.

  2. 2

    Keep the same size, or make it 2× or 4× larger.

  3. 3

    Press Enhance.

  4. 4

    Compare before and after and download.

In detail

What this enhancer does, and doesn't

It runs Swin2SR's real-world model, which was trained to turn degraded photos (blurred, noisy, saved many times as JPEG, forwarded on WhatsApp) into clean, larger ones. The model works at 4×; choose Same size and the cleaned 4× result is scaled back down with high-quality resampling, which leaves a sharper, cleaner photo at the original dimensions.

It does not change colors or lighting, restore faces, fix motion blur from a shaking camera, or colorize black-and-white photos. Heavy blur stays blurry. It works best on photos that are small, noisy or blocky from compression.

How AI upscaling works here

Ordinary resizing stretches pixels, so a small image only gets blurrier. A super-resolution model has learned from millions of pairs of small and large images what fine detail usually looks like, and draws new pixels for edges, text and texture. This tool runs Swin2SR (Apache-2.0, from the University of Würzburg), a transformer model published in 2022, in your browser. Photos go through the real-world model, trained on noisy, blurred and JPEG-compressed pictures, which also cleans those flaws away; graphics, logos and screenshots go through the classical models, which keep clean lines exactly as they are.

The image is processed in overlapping tiles of 192 pixels, so a large picture never needs more memory than one tile at a time, and the tiles are joined without seams. The model is a 54 MB download the first time; your browser keeps it. Nothing is uploaded.

Speed, size limits and honest expectations

  • It is slow on a processor: roughly 3 to 6 seconds per 192-pixel tile, so a 1000 × 750 photo (24 tiles) takes 1 to 3 minutes. With WebGPU it is many times faster. The page shows the estimate before you start.
  • Inputs up to about 1 megapixel for photos and for 4× (a 4× result of that is already 16 megapixels), and 2 megapixels for graphics at 2×.
  • It sharpens what is there; it cannot recover details that are not in the image. Faces in tiny photos will look cleaner, not like the real person, and text too small to read will not become readable.
  • Best for small product shots, screenshots, logos, old web images and thumbnails. Already large, sharp photos gain little.
FAQ

Questions, answered.

Is it free with no watermark?

Yes. It runs on your device, so there are no credits, no watermark and no size-limited preview.

Is my image uploaded?

No. The model runs inside your browser tab; the image stays on your device.

Why is it slow?

Super-resolution is heavy work. Without WebGPU it runs on the processor. Chrome or Edge on a laptop with a recent graphics chip is several times faster.

Can I use the result commercially?

Yes. Swin2SR is under the Apache-2.0 license, which allows commercial use, and the image is yours.

Can it fix a blurry photo?

Mild softness from low resolution and compression, yes. Strong motion blur or a missed focus, no: the detail is not in the photo to recover.

Will it fix old family photos?

It cleans scan noise and sharpens a little. It does not repair tears, scratches or faded colors.

Keep going

More free image tools