AI Image Upscaler

Neural network super resolution that adds realistic detail

How It Works

The upscaler uses Swin2SR, a transformer-based neural network for image super resolution. Ordinary upscaling interpolates between existing pixels, so an enlarged image gets softer the bigger it gets. Swin2SR instead predicts what the extra pixels should look like, based on patterns learned from large sets of real photographs.

The model runs entirely in your browser through transformers.js, on ONNX Runtime's WebAssembly backend. The first run downloads the model (about 7MB), which your browser then caches, so later runs skip the download.

What is Super Resolution? Super resolution upscales an image while adding detail that was not in the original. The network has learned what details typically exist at higher resolutions (skin texture, fabric weave, fine edges) and synthesizes them into the enlarged image. The result looks natural, but the added detail is a prediction, not recovered data.

Experimental Feature: Browser-based AI upscaling does not match native desktop tools like Real-ESRGAN. For critical work, dedicated software will give better results.

Key Features

Learned detail

The neural network generates plausible detail rather than just interpolating pixels.

4x upscaling

Enlarges images 4x in each dimension, 16x the pixel count.

Private processing

The model runs in your browser. Images never leave your device.

Cached model

The model downloads once, about 7MB, and is cached for later runs.

Best Use Cases

Enlarge low-resolution scans of old photographs.
Upscale small images saved from websites.
Turn thumbnails into usable sizes.
Recover some sharpness in heavily compressed images from social platforms.

AI Resize vs Standard Resize

This site has two upscalers, and they suit different jobs. AI Resize predicts new detail with a neural network. It is the right choice when the source is small and you want the result to look like a real photo, accepting that the added detail is invented rather than recovered. The standard Resize tool uses Lanczos interpolation instead: it is much faster, works on any device, and never invents anything, which makes it the safer choice when accuracy matters more than added detail.

Technical Details

The model is Xenova/swin2SR-realworld-sr-x4-64-bsrgan-psnr, a Swin2SR variant trained for real-world photos, loaded with Hugging Face's transformers.js library. Inference runs on ONNX Runtime's WebAssembly backend, so it works in any modern browser without needing a GPU. The download is about 7MB and the browser caches it after the first run.

Processing Time

How long a run takes depends on your device and the size of the source image. Small images finish in seconds; larger ones can take minutes, and very large ones may run out of memory on limited devices. The model quadruples each dimension, so the work grows quickly with input size. Starting from a smaller source is the most reliable way to keep processing time reasonable.

Tips for Best Results

Start from a reasonably sharp source. The network cannot undo motion blur.
Smaller source images process faster and often upscale more cleanly.
Faces and natural scenes tend to upscale well.
Text and sharp graphics can show artifacts. The standard resize tool handles those better.

Limitations

Browser-based inference is slower than a native application.
Very large images can exceed the browser's memory limits.
Results are good but not on par with desktop tools like Real-ESRGAN.
The added detail is predicted by the network, not recovered from the original.

Try AI Upscaler →