Image Tools

Enlarge a small image with a neural network

Add a small image and it is enlarged three times by a neural network rather than by stretching. It runs on your machine, so nothing is uploaded — and it is worth knowing up front what this can and cannot do.

Add a small image and it is enlarged 3× by a neural network rather than by stretching.

The model runs in your browser. Your image is not uploaded.

  • Runs in your browser
  • No account required
  • Free to use
  • Up to 50 files

How it works

01

Sharpening brightness, not inventing colour

The network is a sub-pixel convolution model from the ONNX model zoo, and it was trained on the brightness channel alone. So brightness detail is reconstructed by the model and colour is carried up by ordinary interpolation. That is a real constraint of this architecture, and it is why the tool describes itself this way rather than as "AI enhancement".

02

It works in tiles, and hides the seams

The model takes a fixed 224×224 patch, so a photograph is cut into overlapping tiles and reassembled. Only the middle of each tile is kept: a convolution has least context at the edge of its input, and keeping those edges is what makes a tiled upscale look like a grid.

03

There is a size limit, and a reason for it

Above about 1200×1200 the tile count grows past what is reasonable to run in a phone or laptop tab, so it is refused rather than left to freeze. Shrink the image first, or use the plain resizer if you only need it larger on screen.

What it does

A genuine model, running locally

Small — 240 KB of weights on top of the runtime — but a real network, not a sharpening filter wearing the word "AI".

Fixed 3×

That is what this architecture produces. A slider offering arbitrary factors would just be resampling with extra steps.

Before and after together

The comparison is the only honest way to judge an upscaler, so both are on screen.

Nothing uploaded

The runtime and the model come from this site and the work happens in your browser.

Practical advice

Getting a better result

  • This recovers edge detail. It cannot recover a face that is twelve pixels across, and no upscaler can — the information is not in the file.
  • It does best on clean sources: line art, logos, screenshots and photographs that were downscaled rather than heavily compressed.
  • A heavily compressed JPG will have its artefacts sharpened along with everything else. Start from the best version of the image you have.
  • For making an image larger on screen without changing the file, the resizer is the right tool and it is instant.
  • If you need more than 3×, run the result through the resizer afterwards rather than upscaling twice — a second pass sharpens the first pass’s guesses.

Questions

Things people ask

Is this the same as the AI upscalers I have seen?

It is the same idea and a much smaller model. The impressive ones run networks of tens or hundreds of megabytes on a server GPU. This is 240 KB of weights on your own processor, and it is described plainly for that reason.

Can it recover detail that is not there?

No, and nothing can. An upscaler predicts what a plausible higher-resolution version would look like. It does not recover the original scene, which is why a blurred number plate stays unreadable.

Why only 3 times?

That is what this network was trained to produce. Offering other factors would mean resampling the result afterwards, which is the resizer’s job and is better done there explicitly.

Why is there a size limit?

The work is done tile by tile on your own processor. A large photograph would mean hundreds of tiles and a frozen tab, so it is refused with an explanation instead.

Is my image uploaded?

No. The model runs in your browser and the image never leaves your device.

Who builds these

Free tools by day. Websites that earn their keep by trade.

KuConvert is built by KUDUY, a web design and development agency. From custom WordPress builds to SEO and e-commerce, everything they do is engineered to bring you more qualified traffic and turn it into revenue.

  • Fixed scope, no surprise invoices

    Most sites ship in four to eight weeks with the scope and the price agreed up front.

  • SEO from day one

    Search is designed into the build — structure, schema and Core Web Vitals — not retrofitted.

  • A team, not a ticket queue

    A project manager, a designer and a developer you can actually reach.

Need your whole site to load this fast?

Images are usually the heaviest thing on a page and the easiest win. KUDUY works on site speed, image delivery and Core Web Vitals for teams who would rather not do it by hand.

This tool runs entirely in your browser. Your files are never uploaded.