Runs offline after the first load

Upscale an image 4x

Drop in a picture that is too small. It is enlarged on your own device by a super-resolution model — the file never leaves the browser tab, and nothing is stamped on the result.

Upscaler

No image handy? Try
Model
Enlarge by
Save as

What actually happens

  1. 01

    The model comes to you

    A 4.6 MB (or 51 MB) ONNX file is fetched once from the model host and kept in the browser cache. Open the page again tomorrow and there is nothing to download.

  2. 02

    Your picture stays put

    The image is decoded into a canvas in this tab and handed to a worker thread. There is no upload endpoint on this site — open the network panel and watch.

  3. 03

    Pieces, then one image

    Big pictures are split into overlapping pieces so the GPU or the WebAssembly heap can hold each one. The overlaps are cross-faded back together, which is why there is no grid in the result.

The two models, and when each one wins

Both are 4× super-resolution networks with licences that allow commercial use, and both are bundled with their licence text. The difference is texture versus time.

Quick — Real-ESRGAN general x4v3

CoderViking/realesr-general-x4v3-onnx

Download

A 34-layer convolutional network, small enough to download in about a second. On a 512×384 test image it finished in 1.0 s on WebGPU and 11.8 s on a single CPU thread; a full megapixel takes 4.6 s and 63 s respectively. Edges, lettering and flat colour come out clean.

It smooths fine texture a little — on a close-up photograph, individual hairs come back softer than the detailed model renders them.

Detailed — Swin2SR real-world ×4

onnx-community/swin2SR-realworld-sr-x4-64-bsrgan-psnr-ONNX

Download
Licence
Apache-2.0

A transformer trained on heavily degraded photos, so it removes JPEG blocking and sensor noise while it enlarges. Measured against the true full-size original of the test image, it scores 32.4 dB against bicubic's 31.2 dB — 1.3 dB of real detail, not sharpening.

Slow, and honestly slow: 8.4 s on WebGPU and 2 min 40 s on the CPU for that same small test image. A full megapixel is 36 s on WebGPU and about 16 minutes on the CPU, where WebAssembly gets one thread.

Read the full licence text shipped with this page

Questions people actually ask

Does my picture get uploaded?

No. The only things this page fetches over the network are the page itself and the model file. Your image is read with the browser's own image decoder, processed in a worker thread in this tab, and written back to a canvas. If you load the page once and then go offline, the upscaler still works.

How is this different from Bigjpg, Upscale.media or Pixelcut?

Those all work by sending your file to their servers, which is why they meter you: Bigjpg's free tier stops at 3000×3000 and 5 MB, ImgUpscaler gives ten images a month. Here the computation happens on your own hardware, so there is no per-image cost to pass on — no account, no queue, no watermark, no cap on how many images you run.

How large can the output be?

Four times the width and four times the height of what you put in. A 1024×1024 input becomes 4096×4096; the 512×384 sample on this page becomes 2048×1536. Pick 2× instead and the same result is resampled down to half of that, which is usually what you want for print or for a web banner.

Why does the free tier stop at one megapixel?

Because 4× of one megapixel is already a 16-megapixel canvas — 64 MB of pixel data before the model has produced anything. Beyond that, a laptop with 8 GB of RAM starts failing halfway through with an allocation error, which is a worse experience than being told the limit up front. On phones the ceiling is half a megapixel for the same reason.

Which of the two models should I pick?

Start with Quick. Switch to Detailed when the source is a real photograph that is noisy, soft or visibly compressed, and you care more about texture than about waiting. On illustrations, logos, screenshots and diagrams the two are close enough that the extra 51 MB download is not worth it.

Can it sharpen a blurry face?

Not really, and this page will not pretend otherwise. Dedicated face restorers such as GFPGAN and CodeFormer produce far better eyes and mouths, but their licences forbid commercial use or derive from weights that do, so they are not shipped here. What you get instead is a general 4× enlargement that treats a face like any other part of the picture.

Does it work on a phone?

Yes, with the Quick model and inputs up to half a megapixel — that model is 4.6 MB, so it arrives quickly even on mobile data. Speed depends on whether the phone's browser offers WebGPU; without it the work runs on one CPU thread and takes a good while longer. The Detailed model is 51 MB and slow enough on phone hardware that the page marks it as a desktop choice rather than blocking it.