✨ AI Image Upscaler

Upscale images 2x or 4x with a real AI super-resolution model — runs 100% in your browser, no upload

✓ Free Forever · ✓ No Signup · ✓ No Upload · ✓ Works Offline

🔒 Your image never leaves your device. All processing happens locally in your browser using Real-ESRGAN (realesr-general-x4v3) via onnxruntime-web — WebGPU with CPU/WASM fallback. No upload, no server, no tracking.

Features

  • Real AI Super-Resolution — Runs Real-ESRGAN, a genuine super-resolution neural network — it reconstructs detail as it enlarges, not just smooth interpolation.
  • 100% In Your Browser — Your image never leaves your device. The model runs locally; only a ~5MB model file downloads once on first use.
  • WebGPU Accelerated — Uses your GPU via WebGPU where supported for speed, and falls back to CPU (WASM) everywhere else — slower, but it still works.
  • Tiled Processing — Big images are split into tiles with a live progress bar, so even large photos won't freeze or crash the tab.

How It Works

  1. Upload image — Drop or select the image you want to enhance and enlarge.
  2. Pick 2x or 4x — Choose your output size, then click Upscale. On first use a ~5MB AI model downloads once.
  3. Download PNG — Compare before and after, then save the enhanced image as a PNG.

Use Cases

  • Enhance old or low-resolution photos
  • Sharpen product images for stores and listings
  • Upscale game textures, art, and avatars
  • Prepare small images for large, crisp prints

Frequently Asked Questions

Is this real AI upscaling?

Yes. It runs Real-ESRGAN (the realesr-general-x4v3 model), a real super-resolution neural network, right in your browser — unlike the classic Upscale Image tool, which uses bilinear interpolation.

Does my image get uploaded?

No. The model runs entirely on your device. The only thing downloaded is the AI model file itself (~5MB, once) from ToolKoala — your image is never sent anywhere.

Why is the first run slow?

On first use the browser downloads the ~5MB model and initializes it. If your browser supports WebGPU it runs on your GPU (fast); otherwise it falls back to CPU/WASM, which is slower — larger images can take a while.

What size images can it handle?

Very large images are scaled down before upscaling — about 2048px on the longest edge on CPU, up to 4096px with WebGPU, and never more than about 4 megapixels — so the 4x result (up to ~67 MP) still fits in your browser. If your browser has a lower limit (iPhone Safari does), it goes smaller automatically. The model outputs 4x; choose 2x for a smaller file — switching after the run is instant, no second pass.

Does it keep transparency?

Yes. A transparent PNG (a logo, an icon, a sticker) comes back transparent. The AI enlarges the colours; the transparency itself is scaled smoothly to match, so edges stay clean instead of turning into a black box.

Notes from Milo 🐨

2026-10-02

I ran a transparent logo through here and got it back on a solid black square: the AI only looks at colour, so the transparency was simply thrown away. Now a transparent PNG comes back transparent, with clean edges. I also found that picking 2x after a run just relabelled the 4x picture, so you downloaded a file called "2x" that was twice the size you asked for. Switching now really makes a 2x (or 4x) image from the result, instantly, with no second pass. There's a Cancel button while it works (on a computer without WebGPU a big photo can take a few minutes), dropping something that isn't an image tells you so instead of doing nothing, and phones with a smaller canvas limit get a picture that fits instead of an error at the very end.

From the blog

  • 📖 How to Upscale an Image with AI (Free, No Upload) — Make a picture higher resolution and clearer with a real AI super-resolution model that runs 100% in your browser. How AI upscaling differs from resizing and classic upscaling, and when each one is the right pick.

Related Tools

💬 Bug, idea, or just to say hi? Tell Milo 🐨 · [email protected]