How is AI upscaling different from just resizing an image?

Ordinary resizing interpolates between the pixels you already have, which enlarges the blur along with everything else. A super-resolution network was trained on millions of image pairs, so it predicts plausible edge and texture detail that interpolation cannot invent, producing a genuinely sharper result.

Is my photo uploaded to a server during upscaling?

No. Both engines run on your own hardware through ONNX Runtime Web or TensorFlow.js, using WebGPU or WebGL acceleration where available. Your image is read into browser memory and no copy is ever sent anywhere.

What is the maximum image size I can upscale?

Input is capped at 1600 by 1600 pixels on the default engine. That guard exists because 4× magnification of a 1600 pixel square already produces a 6400 by 6400 output, around 41 megapixels, which is close to what a browser tab can hold without crashing.

Why does the first run take so much longer than later ones?

The default engine downloads roughly 64 MB of model weights the first time you use it, split into four parts, then your browser caches them. Later runs skip the download and start almost immediately. Switching engines fetches that engine’s own weights once.

When should I choose Fast Mode over the default engine?

Fast Mode uses a much lighter network, so pick it for quick previews, older phones and tablets, or small web graphics where the extra fidelity would not show. It accepts files up to 25 MB. For final output you want to print or publish, the default engine is noticeably better.

How does the 2× option relate to the 4× model?

The network itself always works at 4×. Choosing 2× runs that same pass and then downsamples the result with high-quality smoothing, which keeps the reconstructed detail and often looks cleaner than a native 2× enlargement of the original.

Can I fix a blurry or heavily compressed JPG this way?

Only partly. Upscaling reconstructs detail from what is present, so it magnifies JPG blocking, banding, and noise along with the subject. A clean, small original upscales far better than a large but badly compressed one. Very soft or out-of-focus photos stay soft.

What formats are supported, and is there a watermark?

Bring JPG, PNG, or WebP, and you get a PNG back so no new compression is applied to the enlarged result. There is no account, no daily cap, and nothing is stamped onto the output.

AI Image Upscaler (4×)

Enlarge photos 4× with AI engine algorithm running locally in your browser — no uploads, no watermarks, complete data pr

Files never leave your deviceInstant — no upload waitFree, no sign-up

Deep AI (Default)Highest fidelity. Bigger one-time algorithm download (~67MB), best final quality.Fast ModeLighter and quicker — small algorithm, great for previews and slower devices.

Upscale Multiplier

Choose an image or drag & drop it hereJPG, PNG (transparent), or WebP up to 1600×1600 px · AI runs in browser (GPU accelerated when available)

or import from

An AI image upscaler enlarges a photo by reconstructing detail rather than stretching pixels. Choose an engine and a 2× or 4× multiplier, drop in an image up to 1600 by 1600 pixels, and the model runs on your own device to produce a sharper, larger PNG. Nothing is uploaded at any point.

Why interpolation cannot do this job

Enlarging an image in a normal editor uses bicubic or Lanczos interpolation, which averages neighbouring pixels to fill the gaps. Averaging is mathematically safe and visually disappointing: edges go soft, fine texture turns to mush, and small text becomes unreadable. A super-resolution network takes a different approach. Trained on huge numbers of matched low- and high-resolution pairs, it has learned what a sharp edge, a fabric weave, or a brick face actually looks like, so it can synthesise detail that was never in the small file. The result is a prediction rather than recovered truth, which is why it looks convincing on texture and can go slightly wrong on faces at extreme magnification.

Two engines, and when each is the right pick

Deep AI (default). Real-ESRGAN x4plus weights, about 64 MB downloaded once. It processes your image in overlapping tiles to stay inside browser memory, so a large input takes longer but does not crash. This is the one to use for anything you will print or publish.
Fast Mode. A compact ESRGAN-slim network that loads in seconds and finishes quickly, accepting files up to 25 MB. Ideal for previews, weaker devices, and small web images where the difference would not be visible anyway.

Picking one engine never downloads the other's weights, so trying Fast Mode first costs you nothing if you later switch to the default.

What upscaling handles well, and what it does not

Realistic expectations by source image type.
Source image	Expected result
Small but sharp product photo	Excellent — clean edges and readable detail
Anime, line art, or flat illustration	Excellent — crisp lines, no ringing
Old low-resolution digital camera photo	Good — noticeably sharper, some texture invented
Cropped thumbnail from a larger image	Good, provided the crop itself is in focus
Heavily compressed JPG	Mixed — blocking and banding are magnified too
Motion-blurred or out-of-focus shot	Poor — blur is not something the model can undo
Tiny text below a few pixels tall	Poor — characters may be reconstructed wrongly

Privacy, and why cloud upscalers meter you

Server-side upscalers charge per image because GPU time genuinely costs them money, which is why free tiers cap resolution, add watermarks, or ration credits. Running the model in your browser moves that cost to hardware you already own, which is what makes unlimited full-resolution output possible here. It also settles the privacy question outright: family photographs, unreleased design assets, and client images under an NDA never leave your machine, so there is no third-party copy to worry about.

Getting the best result

Start from the sharpest, least compressed original you can find, not a re-saved copy.
Crop to what you actually need before upscaling. Fewer input pixels means faster processing and headroom under the 1600 pixel cap.
Try 2× before 4×. Modest magnification looks natural more often than maximum magnification.
Compare Fast Mode and the default engine on the same image; the difference varies a lot by subject.
Leave the tab in the foreground on long jobs, since background tabs get throttled.

Where this fits with the other image tools

Upscaled PNGs are big, so pass the result through the image compressor before putting it on a page. If you only need specific pixel dimensions rather than reconstructed detail, theimage resizer is the simpler tool. To isolate an upscaled subject for a listing or thumbnail, use the background remover, and to clear a distracting object out of the frame first, try the object remover.

Frequently Asked Questions

How is AI upscaling different from just resizing an image?
Ordinary resizing interpolates between the pixels you already have, which enlarges the blur along with everything else. A super-resolution network was trained on millions of image pairs, so it predicts plausible edge and texture detail that interpolation cannot invent, producing a genuinely sharper result.
Is my photo uploaded to a server during upscaling?
No. Both engines run on your own hardware through ONNX Runtime Web or TensorFlow.js, using WebGPU or WebGL acceleration where available. Your image is read into browser memory and no copy is ever sent anywhere.
What is the maximum image size I can upscale?
Input is capped at 1600 by 1600 pixels on the default engine. That guard exists because 4× magnification of a 1600 pixel square already produces a 6400 by 6400 output, around 41 megapixels, which is close to what a browser tab can hold without crashing.
Why does the first run take so much longer than later ones?
The default engine downloads roughly 64 MB of model weights the first time you use it, split into four parts, then your browser caches them. Later runs skip the download and start almost immediately. Switching engines fetches that engine’s own weights once.
When should I choose Fast Mode over the default engine?
Fast Mode uses a much lighter network, so pick it for quick previews, older phones and tablets, or small web graphics where the extra fidelity would not show. It accepts files up to 25 MB. For final output you want to print or publish, the default engine is noticeably better.
How does the 2× option relate to the 4× model?
The network itself always works at 4×. Choosing 2× runs that same pass and then downsamples the result with high-quality smoothing, which keeps the reconstructed detail and often looks cleaner than a native 2× enlargement of the original.
Can I fix a blurry or heavily compressed JPG this way?
Only partly. Upscaling reconstructs detail from what is present, so it magnifies JPG blocking, banding, and noise along with the subject. A clean, small original upscales far better than a large but badly compressed one. Very soft or out-of-focus photos stay soft.
What formats are supported, and is there a watermark?
Bring JPG, PNG, or WebP, and you get a PNG back so no new compression is applied to the enlarged result. There is no account, no daily cap, and nothing is stamped onto the output.