Enlarge a photo up to 4× with a neural network running on your own GPU, in your browser. The tool measures your image first, so a drawing gets the illustration model and a compressed JPEG gets its artefacts cleaned instead of sharpened, rather than one model being applied to everything. Nothing is uploaded at any tier.
How it works
- Drop a photo. It is measured on the spot — photo, artwork or compressed — and the panel says which model that routes it to.
- Pick a size and a quality tier. Fast is a resampling filter, not AI; Better and Best are trained models, downloaded once and then cached for offline use.
- Press Upscale, compare the result against the original with the slider, and download it. The result is held, so nothing runs twice.
Questions
- Which tier should I use for a photo?
- Best, for anything you care about: Real-ESRGAN compact, tens of seconds per photo on a typical laptop GPU, and a genuine improvement over any filter rather than a sharpening trick. Better is the one to reach for on a long run or on video, where it is fast enough to be practical. Fast is a Lanczos filter and not AI at all — it is there for an instant preview and for machines without WebGPU, and it adds pixels rather than detail.
- Why did it pick a different model for my image?
- Because one model is wrong for two common inputs, in ways that are worse than merely soft. A photo-trained generator given a drawing invents film-grain texture inside flat colour fills and roughens clean ink lines, so line art is routed to a model trained on illustrations. And any sharpening model given a heavily compressed JPEG sharpens the 8-pixel block grid along with the picture, turning artefacts into permanent geometry, so a compressed source is routed to a denoising model first. The panel names the choice and you can overrule it.
- Does upscaling invent detail that was not there?
- Yes, and that is the honest description of what every AI upscaler does. The model produces texture that is plausible for the surrounding pixels, not the texture the original scene had. That is excellent for photographs and restorations and it is why upscaled output is not evidence: a licence plate that was six pixels wide does not become readable, it becomes a confident guess. Hold the \ key to compare against a plain stretch and see exactly what was added.
- Does my file get uploaded?
- No. The models and all processing run inside your browser on your own device. You can verify this yourself: open your browser’s DevTools, watch the Network tab, and process a file: there are no uploads. After the first visit most tools also work fully offline.
- What do the quality tiers mean?
- Fast is Lanczos resampling — not AI at all, instant, and nothing to download. Better is a tiny trained network (~14 KB) that runs as a shader, so it is quick enough for video. Best is Real-ESRGAN compact, a 5 MB model that takes tens of seconds per image and is the right answer for most photos, and the only AI tier that is practical on video. The model downloads once and is then cached for offline use.
- Does it work on old or blurry photos?
- Upscaling sharpens edges and adds plausible detail, and a source the tool measures as compressed is routed to a denoising model so its JPEG blocks are cleaned rather than sharpened. It cannot recover detail that was never captured, so faces enlarged from very small sizes will look smoothed, not restored.