How to use AI image upscaler
- Choose a photo: a small product shot, an old scan, a thumbnail or a screenshot.
- Pick 2× or 4×. The first time, the tool downloads its AI model (about 5 MB) once; after that it starts straight away, even offline.
- Wait while the model enlarges the photo part by part on your device: seconds with a graphics card, up to a few minutes for a large photo on the processor alone.
- Check the result and download the PNG.
Bigger photos without uploading them
A thumbnail that has to fill a banner, a product shot from an old site, a scan that came out small: stretching a photo in an image editor only spreads the same pixels further apart, so edges go soft and blocky. This tool enlarges the photo with an AI model that redraws it at the new size, adding plausible fine detail such as crisp edges, fabric texture and clean lines, instead of blur.
Most online upscalers send your photo to a server and return the result. This one does the work in your browser. The photo is read, enlarged and saved on your device, so pictures of people, unreleased products and private documents never pass through anyone else's computer, ours included.
How it works
The tool uses Real-ESRGAN's compact general model (realesr-general-x4v3), published by its authors under the BSD 3-Clause licence, run by ONNX Runtime Web in a background worker in this tab. The model always enlarges four times. For 2×, each block of two by two of its pixels is averaged, which keeps the sharper detail the model drew while halving the size.
Large photos are processed in tiles, each with a margin borrowed from its neighbours so the seams don't show, and the progress bar counts the tiles. Where the browser offers WebGPU, the model runs on the graphics card, which in our tests was about twenty-five times faster than the processor (5 seconds instead of over 2 minutes for a 1-megapixel photo); otherwise it runs on the processor in WebAssembly, which is slower but works everywhere.
Transparent PNGs keep their transparency: the alpha channel is enlarged smoothly and applied to the upscaled colours. The result is always a PNG, so no detail is lost to compression before you decide what to do with it; the image compressor can make it smaller afterwards.
What to expect
Upscaling works best on photos that are small or slightly soft: product photos, portraits, landscapes, illustrations and screenshots. It cleans up mild JPEG artefacts and noise along the way. It can't recover information that was never captured, so very small faces and text only a few pixels high come out smoother rather than readable.
To keep the result within the browser's memory, the output is capped at about 16 megapixels at 4× and 8 megapixels at 2×. A larger photo is scaled down to fit before it is enlarged, and the page tells you when that happened.
Questions
- Is my photo uploaded?
- No. The AI model is downloaded to your browser, and the photo is enlarged on your device. The only download is the model itself, from WeaveAI's own engine server; your photo never leaves your device.
- How big is the download, and why is the first run slower?
- The model is about 5 MB and downloads once; browsers that can use the graphics card also fetch a GPU runtime of about 5 MB. Both are then kept by your browser, so later runs start at once.
- How large can the photo be?
- At 4× the result is up to 16 megapixels (for example 4096 by 4096), so photos up to about 1 megapixel are enlarged as they are. At 2× photos up to 2 megapixels are used as they are. Larger photos are scaled down to fit first, and the page says so.
- Does it work on any photo?
- It is trained on everyday photos and does best on soft, small or slightly noisy images. It can't recover detail that isn't there at all: tiny text and faces a few pixels wide come out smoother, not readable.