Real AI upscaling — not just stretched pixels
Every image editor can "enlarge" a photo. What they do is interpolation: stretching the existing pixels and guessing what's between them. The result is always the same — a bigger, blurrier version of what you had. AI upscaling is a different operation. Real-ESRGAN, the open-source model this tool runs, was trained on millions of image pairs to reconstruct plausible fine detail: crisp edges, real textures, readable text. A 4x AI upscale doesn't just have more pixels; it has more information.
And it runs entirely on your device. The model is a small ~4.7MB file that downloads once from a pinned HuggingFace release and caches in your browser. After that, your photo is processed tile by tile in your tab via WebAssembly — never uploaded, never stored, never seen by anyone. You can watch the network tab go quiet and stay quiet.
The honest trade-off is time. There's no free GPU farm here; your CPU does the work. A full photo takes roughly 1–3 minutes on a modern desktop, longer on a phone. The tool shows per-tile progress so you always know where it stands, and the input is capped at a 1024px long side to keep memory and time sane on ordinary devices. What you get for the wait is a genuinely sharper image — the kind of result that used to require paid desktop software.
2x or 4x — which to pick?
4x quadruples each dimension — a 1024px photo becomes 4096px, sixteen times the pixels. That's the choice for prints, posters, and rescuing old low-res photos you want to actually use. 2x gives you the same AI-enhanced reconstruction resampled to half that size: still dramatically sharper than any normal 2x resize, in a smaller file that's friendlier for screens, slides, and uploads with size limits. You can switch after upscaling too — the 2x is derived from the full 4x result, so there's no quality penalty for changing your mind.
How it works
- Pick 2x or 4x. Choose your target before uploading — you can change it after, too.
- Choose a photo. Tap the big button or drop an image in. JPG, PNG, WebP, HEIC (iPhone), AVIF, BMP, GIF, SVG and TIFF all work — read locally, never uploaded.
- The AI upscales in tiles. First visit downloads the ~4.7MB model (cached after). Your photo is processed in overlapping 256px tiles so it fits in memory; seams are cropped away, so the result is seamless.
- Download PNG or JPG. Full AI resolution, no watermark.
Where AI upscaling earns its keep
Old photos worth saving. The classic: a 640px photo from 2009 that you want printed, framed, or just viewable on a modern screen without turning to mush. AI upscaling is the single biggest quality jump available for these.
Product and real-estate shots. A slightly soft listing photo upscaled 2x looks noticeably more professional in a thumbnail grid — and thumbnails are where the click happens.
Screenshots and graphics for print. A 2x upscale of a UI screenshot or infographic gives print layouts the pixel density they need without the smeared look of plain resizing.
AI-generated images. Many generators output at modest resolutions. A 4x pass turns a 1024px generation into a 4096px file you can actually print or use as a wallpaper.
Video stills and frames. A grabbed frame is rarely more than 2MP. Upscaling 4x makes it usable as a thumbnail or a poster background where the plain frame would fall apart.
Tips for the sharpest result
- Start with the best source you have. The AI reconstructs detail from what it can see — a clean original upscales beautifully; a photo that's already been through three rounds of WhatsApp compression has less to work with.
- Don't pre-sharpen. Sharpening halos and heavy JPEG artifacts get faithfully "enhanced" too. Upscale first, sharpen lightly after if needed.
- Match the scale to the job. 4x for print, 2x for screens — bigger isn't always better, it's just bigger.
- Keep the tab in front. Browsers throttle background tabs; on phones especially, leaving the tab visible keeps the upscale running at full speed.
- Compare at 100%. The side-by-side preview shows the difference, but zoom to full size on edges and text — that's where the AI earns its keep.
Honest limits
AI upscaling is reconstruction, not magic. It cannot invent information the photo never captured: a heavily blurred face becomes a sharper heavily blurred face, not a new face. Tiny thumbnails (under ~200px) upscale into something cleaner but still visibly limited. Text smaller than a few pixels may come back as plausible-looking but wrong glyphs — always verify small text at 100% before publishing.
Two more honest notes. First, inputs are capped at a 1024px long side before processing, so the 4x output is at most ~4096px — a deliberate guard so the tool works on real devices instead of crashing them. Second, the model is RGB-only, so semi-transparent PNG areas are composited onto white first. For everyday enlarging — old photos, prints, thumbnails, AI art — on-device Real-ESRGAN now matches what the paid hosted upscalers do, without the upload, the account, or the per-image credits. Pick a photo, wait a few minutes, download. That's the whole workflow.