Image Upscaler — Double the Size in Your Browser
Double the size of an image with a super-resolution model that reconstructs edges and texture instead of stretching pixels. It runs in your browser: no account, no credit system, no watermark, and the picture is never uploaded. The model downloads once, about 20 MB, and is then cached.
Ordinary resizing interpolates — it averages neighbouring pixels, which is why an enlarged image looks soft or blocky. A super-resolution network has been trained on millions of pairs of small and large images, so it predicts what the missing detail most likely was: it sharpens edges that interpolation would smear and rebuilds texture that was flattened. The model here is Swin2SR, a transformer-based architecture designed for exactly this task.
What upscaling can and cannot do
It recovers plausible detail, not real detail. Text that has become an illegible smudge will come back as sharper nonsense, a face reduced to twenty pixels will not become a recognisable person, and no model can undo heavy JPEG artefacts perfectly. What it does well is take an image that is merely small — a product photo from an old listing, a logo from a website, a scanned photograph, a screenshot you need at twice the size — and make it usable at a larger size without the softness of a plain resize.
Why there is a size limit
The model processes every pixel on your own processor, and the work grows with the area, not the width: doubling the input side quadruples the time. Inputs are therefore scaled to 512 pixels on the longest side, which still produces a 1024-pixel result and keeps the wait to seconds or a minute or two rather than an unpredictable stall. Hosted services do this on server GPUs, which is why they can offer 4x and 8x — and also why they ask you to upload the image, sign in, or pay after three files.
A practical order of operations
Upscale first, then edit. If you plan to crop, do it before upscaling so the model spends its resolution on the part you keep. If the upscaled PNG is larger than you need for the web, run it through the image compressor afterwards — a 2x PNG is often several megabytes. For a straight dimension change with no model involved, the image resizer is faster and exact.
How to use Image Upscaler
- Drop in a PNG, JPG, or WebP. Small images work best — anything over 512 pixels on the longest side is scaled down to that first.
- Wait for the model. The first run downloads about 20 MB; later runs start immediately from the browser cache.
- Let it work. Depending on the device this takes from a few seconds to a couple of minutes, and the tab has to stay open.
- Flip between the original and the result with the compare button to see what the model rebuilt.
- Download the PNG, and compress it afterwards if it is destined for a web page.
Features
- Real super-resolution — the Swin2SR model rebuilds edges and texture rather than interpolating between pixels.
- 2x output — a 500-pixel image comes back at 1000 pixels.
- No credits, no watermark, no sign-up — the common limits of hosted upscalers do not apply because there is no server involved.
- Before-and-after compare — switch between the original and the result at the same display size.
- Cached after the first run — the model stays in the browser, so later images start straight away and work offline.
- The image never leaves your device — nothing is uploaded, stored, or logged.
Frequently Asked Questions
How do I enlarge an image without losing quality?
Use a model rather than a resize. Drop the image in above and it comes back at twice the size with edges and texture reconstructed instead of averaged. A plain resize can only interpolate between the pixels that already exist, which is what makes enlarged images look soft.
Is it really free, with no watermark or sign-up?
Yes. The model runs in your browser, so there is no server cost to recover — no credits, no account, no watermark, and no cap on how many images you process. The trade-off is the size limit and the processing time, both of which come from using your own hardware.
Why is there a 512-pixel input limit?
The work scales with the number of pixels, so doubling the input side quadruples the time, and large inputs can exhaust the tab’s memory. Capping the input at 512 pixels keeps a 1024-pixel result within a reasonable wait on an ordinary laptop or phone.
Can it make blurry text readable again?
Only if the text was still almost legible. The model predicts plausible detail from what remains; when the strokes have merged into a smudge, the result is a sharper smudge rather than the original words. The same applies to faces that have been reduced to a handful of pixels.
How long does it take?
From a few seconds on a recent laptop to a couple of minutes on an older phone, plus the one-time model download. The tab must stay open and in the foreground on mobile, since background tabs are suspended.
Why is the downloaded file so large?
The output is PNG, which is lossless, and it has four times the pixel count of the input. For anything going on a web page, run the result through the image compressor and convert it to WebP or JPEG.