Enlarge images 2×, 3× or 4× without the mush of an ordinary resize. The neural network runs entirely in your browser — no upload, no account, no limit on how many images you run.
Drop an image to upscaleor click to browse · JPG, PNG, WebP, AVIF
AI reconstructs detail and takes longer. Fast is a high-quality classical resample.
The AI model works at 4× natively; 2× and 3× are rendered down from it, so every factor costs about the same.
PNG is lossless. JPG and WebP give much smaller files.
Only applies to JPG and WebP output.
92%
—→—
Original
Upscaled
Both previews are shown at the same width so you can compare detail directly — the upscaled file is larger than it looks here. Download it to see the real resolution.
Freeno limits, no watermark
4×up to quadruple resolution
100%private — runs in browser
How AI upscaling differs from resizing
A normal resize has to guess the colour of every new pixel from the ones next to it, which is why enlarged photos look soft and plasticky. A super-resolution network was instead trained on millions of pairs of small and large images, so it has learned what edges, textures and text actually look like at higher resolution and reconstructs them rather than smearing.
QuickImg runs the Swin2SR model through WebAssembly in your own browser. The model weights are downloaded once from a CDN and cached; your picture is never sent anywhere. That is also why there is no daily quota — the work happens on your machine, not on a server we have to pay for.
Getting the best result
Start from the largest original you have. Upscaling an image that was already enlarged once compounds the artefacts.
AI mode shines on photos, faces and product shots. For flat graphics, logos and screenshots, Fast mode is often cleaner.
Big inputs are slow: the model processes the image in tiles, so a 4000 px photo at 4× is thousands of tiles. Try 2× first.
Export to PNG if the result is going into further editing, and to WebP or JPG if it is going straight onto a web page.
Frequently Asked Questions
Is the image really not uploaded anywhere?▸
Correct. The AI model file is downloaded from a CDN to your browser, then all inference runs locally through WebAssembly. Your image data never leaves the device. You can confirm it by opening your browser devtools Network tab — no request carries your file.
Why is the first run slow?▸
The first time you use AI mode the browser downloads the Swin2SR model and the ONNX runtime — about 25 MB in total, or about 55 MB if your browser can run it on the GPU, which uses higher-precision weights and is far faster. It is cached afterwards, so later runs skip straight to processing. Fast mode needs no download at all.
How large an image can I upscale?▸
There is no hard cap, but the practical limit is your device memory. The image is processed in tiles so very large photos work, they just take a long time. Above about 4 megapixels the tool suggests Fast mode instead.
What is the difference between 2×, 3× and 4×?▸
The model reconstructs at 4× natively. 2× and 3× are rendered down from that 4× result, which is why they look cleaner than resizing the original directly — and why every factor takes roughly the same time.
Can it recover detail that is not in the photo?▸
No tool can. A super-resolution model produces a plausible reconstruction based on what it learned during training, which usually looks far better than interpolation, but it is not evidence of what was actually there. Do not rely on it for forensic or legal purposes.