Make a small image bigger

Not for the free result. The model file comes from us once and your browser runs it on your own processor; we count that a run happened and learn nothing about the picture. Only if you press "Better quality · 2 credits" is your original uploaded: it goes to our server and to the AI provider that makes the enlargement (which does not train on it and deletes it within the hour), and the upload and the result are deleted from our servers after 2 hours.

It depends on the size of the picture you start with, and the page tells you before you start. Measured on the model this page ships, on the single-threaded WASM backend that a browser without WebGPU uses: about 12 seconds for a 288×216 image, and 242 seconds — about four minutes — per megapixel of input. A small icon or screenshot is seconds; a 1,000×1,000 photo is about four minutes, whichever size you pick, because the model does the same work on every input pixel. On WebGPU it is roughly twice as fast, not ten times — the model is small and the work comes in tiles, so the GPU spends much of its time waiting to be handed the next one.

About 32 megapixels on a computer — roughly 6,500×4,900 — and about 16 on a phone or a machine with little memory. The limit is memory, not time: the finished picture has to fit in your browser tab, and 32 megapixels is already 128 MB of it. If the size you picked would be larger, the page does not refuse; it makes the picture as large as your device can, such as 2.6× instead of 4×, and says so. Only a picture that is already about that size is turned away, because there is nothing left to enlarge it into. The better version (2 credits, signed in) is different: it runs a much larger model on our server that rebuilds fine texture rather than only sharpening — the one to use on a small photograph — and makes results up to 16 megapixels on any device; the page tells you the exact size before you press it.

Because that is what the model was trained for. realesr-general-x4v3 has a fixed four-times scale; asking a network for a size it was never trained at gives a worse result than running it natively and halving the output. It costs the same time either way.

Small, faint visible text reading "AI-edited · filedone.ai" in the bottom-right corner, plus an XMP metadata packet recording that the image was enhanced by an algorithm, which tool did it, and an id derived from the finished pixels. The browser version carries no cryptographic signature, so the label can be stripped. The better version is made on our server and signed there: the same metadata plus an embedded C2PA signature and an invisible watermark, and no visible text.