← All posts

How to Upscale an Image with AI (Free, No Upload)

2026-07-24

Short answer: Drop your image into the AI Image Upscaler, pick 2x or 4x, and click Upscale. A real neural super-resolution model (Real-ESRGAN) reconstructs detail and hands you back a bigger, cleaner PNG. It runs entirely in your browser — the ~5MB model downloads once, then your photo never leaves your device.

If you've ever blown up a small image and gotten a soft, blocky mess, this is the fix. But "make it bigger" actually means three different things, and picking the wrong one wastes your time. Let me untangle them first, because it's the part every other guide skips.

Resize vs. classic upscale vs. AI upscale

These get lumped together as "upscaling," but they're not the same operation:

  • Resize just changes the dimensions. If you shrink a 4000px photo to 800px, that's a clean downscale. Going the other way — 800px up to 4000px — a plain resize stretches the pixels you have. No new information appears; it just gets bigger and blurrier. Use Resize when you need an image at an exact pixel size and it's already large enough.
  • Classic upscale (bicubic/Lanczos interpolation) is smarter math for filling the gaps between pixels. It smooths the stair-steps so the result looks less jagged than a raw stretch — but it still can't invent detail that isn't in the source. Edges get softer, not sharper. The classic Image Upscaler does this, and it's genuinely the better pick in two cases: an already-sharp image where you just need more pixels, or when you need an exact output dimension. It's also instant — no model to load.
  • AI upscale is the different beast. A neural network trained on millions of image pairs reconstructs plausible detail: it rebuilds edges, textures, and fine structure that interpolation can only blur. That blurry old photo or tiny thumbnail actually gets sharper, not just bigger. The tradeoff is it's heavier (a model has to run) and it's inventing detail — great for photos, occasionally weird on text or fine logos.

So: exact dimensions or an already-crisp source → classic upscale. Something small, soft, or low-res that you want to genuinely rescue → AI.

How to AI-upscale a picture

  1. Open the AI Image Upscaler.
  2. Drop in a JPG, PNG, or WebP (iPhone HEIC works too — it gets converted automatically).
  3. Choose 2x or 4x. The model natively outputs 4x; picking 2x runs the same model and scales the result back down to 2x, which is often the cleaner look for photos.
  4. Click Upscale. The first run downloads the model (~5MB, cached after that), then processes.
  5. Compare before/after side by side and download the PNG.

Why it runs in your browser (and what that costs)

Most "AI upscaler" sites upload your photo to their server, run it there, and often keep a copy. This tool doesn't — the model runs locally with WebGPU or WebAssembly. Your image is never transmitted. You can open DevTools → Network during a run and watch: the only request is the one-time model download, never your file.

The honest tradeoff is speed. If your browser supports WebGPU (recent Chrome/Edge), it's fast — a second or two. On CPU (WebAssembly) it's slower, maybe 10–30 seconds for a normal photo. The tool shows which backend it's on. To keep things responsive it also caps very large inputs to a safe edge size before processing, so a huge source may get lightly downscaled going in — you still get the full upscaled result, just computed sanely.

Tips for better results

  • Start small. AI upscaling shines on genuinely low-res inputs. A 300px face → 1200px is a dramatic rescue. Upscaling an already-large 3000px photo gains little and takes longer.
  • Don't stack it blindly. Running 4x then 4x again rarely helps and can amplify artifacts. One pass at the scale you actually need.
  • Photos over graphics. The model is trained on real-world imagery. Screenshots with sharp text or flat vector logos are better handled by the classic upscaler, or just export the vector larger.
  • Output is PNG. It's lossless, so it preserves the reconstructed detail — but it can be large. If you need it smaller afterward, run it through Compress.

FAQ

How do I upscale an image with AI for free? Use an in-browser tool like the AI Image Upscaler. Drop in your image, pick 2x or 4x, and a real super-resolution model reconstructs detail locally. It's free and nothing gets uploaded — the only network request is a one-time ~5MB model download.

What's the difference between AI upscaling and just resizing? Resizing (or classic interpolation upscaling) stretches or smooths the pixels you already have — it can't add detail, so bigger means blurrier. AI upscaling uses a neural network to reconstruct plausible new detail — edges and textures the source didn't clearly contain — so the result actually looks sharper, not just larger.

Does the AI upscaler upload my photo? No. The model runs entirely in your browser via WebGPU or WebAssembly. Your image stays on your device; you can verify it in DevTools → Network, where the only request is the model file, never your photo.

When should I use the classic upscaler instead? Pick the classic Image Upscaler when your source is already sharp and you just need more pixels, or when you need an exact output dimension. It's instant (no model to load) and won't invent detail you don't want. AI is for rescuing small, soft, or low-resolution images.

Why is it slow on my computer? If your browser doesn't support WebGPU, the model falls back to CPU (WebAssembly), which is noticeably slower — typically 10–30 seconds for a normal photo. Using an up-to-date Chrome or Edge with WebGPU makes it much faster.

— Milo 🐨