Skip to content

Upscale images with AI

Enlarge an image 2× or 4× with an AI super-resolution model running in your browser, keeping detail without uploading the picture.

Drop an image to upscaleor click to browse · paste with Ctrl+V · up to 1.1 MP free
Choose imageProcessed on your device — never uploaded
Runs in your browser · no uploadHow it works

How to use Image upscaler

  1. 1

    Add an image

    JPG, PNG, WebP or HEIC up to about 1.1 megapixels; larger pictures can be shrunk to fit in one click.

  2. 2

    Choose scale and model

    Pick 2× or 4×, then Best quality for the sharpest result or Fast for slower computers.

  3. 3

    Upscale

    Follow the tile-by-tile progress and estimated time left; cancel whenever you like.

  4. 4

    Compare and save

    Inspect details with the slider at 100% zoom, then download PNG, JPG or WebP.

Limits and Pro

Input images up to about 1.1 megapixels (for example 1280 × 860) and 25 MB, upscaled 2× or 4× with no watermark.

Inputs up to 4.2 megapixels at 2× and 2.1 megapixels at 4× — the most a browser tab holds comfortably in memory — and files over 25 MB. See Pro

About Image upscaler

The upscaler returns an image exactly two or four times wider and taller than the one you add, rebuilt by a Swin2SR super-resolution transformer rather than stretched with bicubic interpolation. A 640 × 480 photo becomes 1280 × 960 at 2× or 2560 × 1920 at 4×. The picture is processed in overlapping tiles and only the centre of each tile is kept, so no seams appear; transparency in PNG or WebP files survives because the alpha channel is resampled separately. Save the result as lossless PNG, or as JPG or WebP with adjustable quality; EXIF metadata is not carried over.

Three models are available: a lightweight 2× network that is quick on ordinary laptops, the classical 2× model for the crispest edges on clean images, and a real-world 4× model trained on photos with noise, blur and JPEG blocking, which tidies those artefacts while enlarging. Good candidates are small product photos, old scans, thumbnails needed for print, low-resolution screenshots and digital artwork.

Super-resolution predicts plausible detail. It cannot recover text that was unreadable, faces that were out of focus or motion blur, and it is no forensic enhancement tool. Video upscaling and enlargement beyond 4× in a single pass are not supported.

Privacy

Every tile is enhanced by an ONNX model running in Web Workers inside your browser, so your file never leaves your device and no server ever sees the pixels. The browser fetches the chosen model once (8 to 55 MB from Hugging Face) and keeps it cached, which also lets the upscaler work offline.

Features

Real super-resolution

Swin2SR transformers reconstruct edges and texture instead of interpolating pixels.

Seamless tiling

Overlapping tiles keep memory low on large images without visible joins.

Artefact clean-up

The real-world 4× model removes JPEG blocking and sensor noise while it enlarges.

AI upscaling compared with ordinary resizing

Bilinear and bicubic resizing only interpolate between existing pixels, so an enlarged picture looks soft and blocky. A super-resolution network has learned from millions of low- and high-resolution image pairs what fine structure usually looks like, and predicts plausible edges, textures and lines for the new pixels.

This tool uses three Apache-2.0 Swin2SR checkpoints: the classical ×2 model for clean images, a lightweight ×2 model for speed, and a real-world ×4 model trained on photos degraded by noise, blur and JPEG compression.

Image upscaler: frequently asked questions

How big can the upscaled image be?
Output is the input size multiplied by two or four in each direction. On the free plan the input can be up to about 1.1 megapixels, so the largest free 4× result is roughly 17 megapixels, for instance 5120 × 3440 pixels. Pro raises the input to 4.2 megapixels at 2× and 2.1 megapixels at 4×. Bigger photos can be shrunk to fit before upscaling.
Why is upscaling slow on my laptop?
Swin2SR performs billions of calculations per tile. In Chrome or Edge with WebGPU it runs on the graphics card and finishes in seconds. Without WebGPU it falls back to the processor, where each browser worker runs on a single thread, so the tool spreads the tiles across two or three workers. The Fast model is several times quicker than Best quality on such machines.
What is the difference between 2× and 4×?
2× doubles the width and height, giving four times the pixels, using either the lightweight or the classical model. 4× quadruples each side for sixteen times the pixels; Best quality then uses a model trained on real-world photos that also removes noise and compression artefacts, while Fast runs the lightweight 2× model twice. For most web and print uses, 2× is enough and much quicker.
Will AI upscaling make a blurry photo sharp?
It sharpens edges and restores texture lost when an image was downsized or heavily compressed, which is the most common reason a photo looks soft. It cannot reconstruct detail that was never recorded: an out-of-focus face, strong motion blur or unreadable small print will look cleaner but not truly legible. For noisy phone pictures the real-world 4× model gives the most natural result.
Does upscaling keep transparency and metadata?
Transparency is kept when you save as PNG or WebP: the alpha channel is enlarged with high-quality resampling and recombined with the upscaled colours. JPG has no alpha channel, so transparent areas turn white. Camera metadata such as EXIF tags, GPS position and embedded colour profiles is not copied to the output, and phone photos are rotated upright before processing.

Related tools

All image tools

Background remover

Remove image backgrounds with on-device AI

Image resizer

Resize by pixels, percent or preset — in bulk

Image compressor

Shrink JPG, PNG, WebP and AVIF without visible loss

HEIC, WebP & AVIF converter

Convert HEIC, WebP and AVIF to JPG or PNG — and back