- Is my image uploaded anywhere?
- No. The model is downloaded to your browser once — a 2.5 MB file plus the runtime that executes it — and then the upscaling happens on your own machine, on the graphics card where your browser supports WebGPU and on the processor where it doesn't. Your picture is read from your disk, processed in the tab, and saved back to your disk. You can check it: open the network panel in your browser's developer tools, drop an image in, and watch — the only requests are for the model and runtime files, and none of them carry your photo. That's also why there are no credits here. Upscalers that run the model on their own servers pay for every image you process, so they have to meter you; this one doesn't pay for anything, so it has nothing to meter.
- Does upscaling really add detail, or does it just make the image bigger?
- It adds detail that looks right, which is not the same as detail that was there. A plain resize can only spread the pixels you have across a bigger area, so edges turn into soft ramps. A super-resolution model was trained on millions of pairs of sharp images and deliberately degraded copies, so it has learned what a sharp version of a blurry edge, a strand of hair or a brick wall tends to look like, and it draws that. For a product photo, a thumbnail or an old profile picture the result is usually far better than any resize. But the model is predicting, not recovering: small text comes out as confident squiggles, faces can shift subtly, and a license plate that was unreadable before will still be unreadable — just sharper-looking. Never use an upscaled image as evidence or as an ID photo.
- Why only 2× and 4×? Other sites offer 8× and 16×.
- Because past 4× the model is inventing most of what you see. At 4× every source pixel becomes sixteen output pixels, and the model has enough context from the neighborhood to fill them convincingly. At 8× it's sixty-four pixels per source pixel, at 16× it's two hundred and fifty-six, and the result starts to look painted — smooth, plasticky surfaces and textures that were never in the photo. If you genuinely need a huge print, upscale 4× here, then resize the result with Image Resizer; you'll get the same pixel count without pretending the extra detail is real.
- Why won't it upscale my big photo?
- Browsers cap how large an image they can hold in memory, and the tightest limit — Safari on iPhone and iPad — is 16,777,216 pixels, which is a 4096 × 4096 square. Past it, the browser doesn't fail loudly; it quietly saves an empty image, so this tool refuses instead of handing you a blank file. At 4× that means sources up to about 1024 × 1024; at 2×, up to about 2048 × 2048; and no side of the result can pass 8192 pixels. If only 4× is too big, the tool switches to 2× on its own and tells you. If even 2× is too big, your image is already 4 megapixels or more, or over 4096 pixels on a side — big enough for almost any screen — and it says so rather than wasting your time.
- How long does it take?
- It depends on your hardware more than on the image. With WebGPU — current Chrome and Edge on most laptops and desktops — the model runs on your graphics card, and a typical 800 × 600 photo takes anywhere from a few seconds to half a minute depending on the card. Without it, the model runs on your processor, one tile at a time, at roughly five seconds a tile on a recent laptop, so the same twelve-tile photo takes a minute or more; the progress bar shows how many of the tiles are done so you're never staring at a spinner. The very first image also downloads the model and runtime, which is a one-time wait of a few seconds on a normal connection.
- Does it keep transparency?
- Yes. The model only understands colour, so a transparent PNG or WebP is split in two: the colour is upscaled by the model and the transparency mask is upscaled separately, then the two are put back together. Before the model sees the image, the invisible pixels around a cut-out are filled with the colour of the nearest visible pixel, which is what stops the dark halo you get from tools that treat transparent areas as black. Areas that were fully transparent stay fully transparent in the download.
- What's the best kind of image to upscale?
- Anything that's small but not badly damaged: a product photo that's too small for a marketplace listing, a logo you only have as a small PNG, a screenshot, a thumbnail you need at full width, an old digital-camera photo. The model is also trained to clean up JPEG compression, so photos that have been through messaging apps usually come out cleaner as well as bigger. What it can't do is fix motion blur, heavy noise from a dark photo, or a subject that's out of focus — those carry too little information for any model to work from.
- Why doesn't this charge credits like other upscalers do?
- Most upscalers in this category follow the same pattern: upload your photo, process it on their GPU, then put the full-resolution download behind an account, a watermark, a daily cap or a monthly credit pack. That isn't greed so much as arithmetic — every image costs them server time. Compact super-resolution models now run fast enough inside a browser tab that the server, and the bill that comes with it, can simply be removed. This tool is the model, the tiles, a before-and-after divider and a download button. Nothing else, and nothing to buy.
