Free · On your device · Nothing uploaded

Upscale an Image 4x

A small super-resolution model, running in your browser. It measures your image and tells you how long it will take before it starts.

Drop in a small or soft photo and get a version four times larger with edges and detail rebuilt by a model, not just stretched. The photo never leaves this page, and the line under the upload box counts the bytes to prove it.

Drag, drop or paste · JPG, PNG, WebP · up to about 2,000 px on the long side

4.9 MB model, cached after the first visitEstimates the time firstTransparency kept

Best on photos, screenshots, small logos and thumbnails that need to be printed or shown larger. It rebuilds plausible detail; it cannot recover text that was never legible or a face that was never in focus, and it says so rather than pretending.

What the model actually does

Enlarging an image is easy; every browser can stretch pixels. What a super-resolution model adds is a guess at the detail that a larger version would have had: the crisp edge of a letter, the grain of a fabric, the boundary between hair and sky. This page runs Real-ESRGAN's compact "general x4" model, a 4.9 MB network trained to make that guess well on ordinary photographs, screenshots and graphics. It is the same family of model behind most "AI upscaler" sites, run on your own hardware instead of theirs.

Because it is a guess, it can be wrong in specific ways. Text that was already unreadable becomes confident-looking nonsense rather than words. A face that was blurred stays a blurred face with sharper edges. Heavy JPEG artefacts are sometimes rebuilt as texture. The comparison view exists so you check the result rather than trust it.

Why it estimates the time before committing

The model works in 128-pixel tiles, and the first tile is timed. Multiply that by the number of tiles and you have how long the whole image will take on this device, right now, which is a fact rather than a promise. If the estimate is long, the page says so and offers the instant tier instead of quietly hanging. On a laptop with WebGPU a 1,000-pixel photo takes a few seconds; on a phone without it, the same photo may take a minute, and you deserve to know that before you tap.

What "nothing uploaded" means here, measured

The model weights download once, from this site, and are cached. After that, the only network traffic during processing is zero bytes, and the line beneath the upload box counts them. Open your browser's developer tools on the Network tab while it runs and you will see the same thing. Credit-metered upscalers cannot make this offer, because the upload is how they work.

Instant tier, for when you just need it bigger

The instant option is a stepwise high-quality resample, each step at most double, followed by a light unsharp mask. It is what a good image editor does when you ask it to resize, and it finishes in milliseconds. It is not a model and this page does not call it one. Use it for a quick preview, for very large images, or when the model's estimate is longer than you have.

Common questions

Is my image uploaded anywhere?
No. The model runs in your browser using WebAssembly or WebGPU. The only download is the 4.9 MB model itself, from this site, once. While your image is being processed the page counts the bytes sent to any server and shows the total, which is zero. You can confirm it in the browser's Network panel.
How long does it take?
It depends on your device and the image size, and the page measures it rather than guessing: the first 128-pixel tile is timed and the estimate for the rest is shown before it continues. As a rough guide from our own measurements: one CPU thread takes about 2 seconds per 128-pixel tile, and eight threads (which this page enables in browsers that support cross-origin isolation, so Chrome, Edge and Firefox) bring a 512x384 photo to under 6 seconds. WebGPU is attempted first where the browser offers it and falls back automatically if it fails; we do not quote a speed for it because it varies by device.
Is this the same as the AI upscalers that charge credits?
It is the same kind of model, a Real-ESRGAN, doing the same job. The difference is where it runs. Those sites upload your photo to their servers and charge per image; this one runs the model on your device, so there is nothing to meter and nothing to upload.
Can it fix a blurry photo or make small text readable?
Only partly. It sharpens edges and rebuilds plausible texture, so a soft photo looks crisper. It cannot invent information that was never captured: badly out-of-focus faces stay out of focus, and text that was illegible becomes sharp shapes that are still not the right letters. Check the result at 100% before relying on it.
Does it keep transparency?
Yes. The model itself has no alpha channel, so the page upscales the original transparency separately and reapplies it, and the download is a PNG with real alpha.
What is the largest image it will take?
Inputs up to about 2,000 pixels on the long side, which gives an 8,000-pixel result. Above that the output canvas alone is over 250 MB in memory and most browsers refuse it. Crop or resize first if you need a section of something larger.

Continue crafting

Keep going without re-uploading — your image carries over to the next tool automatically.

  1. 1EditAdd text, effects, or a fresh background
  2. 2CompressShrink the file for fast loading
  3. 3ResizeResize for any platform or social preset