How this AI upscaler works
A normal resize interpolates between the pixels you already have, so a small photo blown up 4× just looks like a blurrier version of itself. This tool instead runs the image through Real-ESRGAN, a super-resolution network trained on millions of degraded / clean image pairs. It has learned what sharp edges, skin, fabric, foliage and text should look like at higher resolution, so it reconstructs plausible detail rather than smoothing everything out.
The model runs through ONNX Runtime Web on WebGPU, using your graphics card directly from the browser tab. If your browser or GPU doesn't support WebGPU, it falls back to a multi-threaded WASM build on the CPU — slower, but the output is identical.
Large images are processed in overlapping 256×256 tiles with a 32-pixel context ring. Each tile sees a little of its neighbours before the border is cropped away, which keeps memory bounded and leaves no visible seams in the assembled result.
What happens to your image
Nothing is uploaded. The file is decoded in a Web Worker, converted to a tensor, run through the model and re-encoded as PNG — all inside the tab. The only network request the page makes is downloading the model weights (~69 MB) the first time you use a given scale; after that they are stored in the browser's Cache API and reused offline.
When to use it
Old and scanned photos
Prints scanned at low DPI, or snapshots saved years ago at a tiny resolution, rarely hold up when printed or framed larger. Upscaling restores enough detail to make the bigger size usable.
Product and listing images
A supplier photo that's too small for a modern storefront template looks soft the moment it's stretched. 2× is usually enough to make it crisp again.
Avatars and profile pictures
A picture that looks fine as a 64-pixel thumbnail falls apart when a site renders it at 400 pixels. Upscale first, then upload.
Re-shared and compressed files
Messengers and social platforms recompress images every time they're forwarded. This rebuilds sharpness lost to that compression.
Art, sketches and illustrations
Small scans of drawings, logos rasterized too small, or thumbnail-sized artwork all gain usable resolution.
Wallpapers and backgrounds
Turn a small source image into something that still looks sharp at desktop or phone wallpaper size.
How to upscale an image
- Pick 2× or 4×. Use 2× for photos that are already reasonably large, 4× for genuinely small sources.
- Drop your PNG, JPG or WEBP onto the panel, or click to browse.
- Press Upscale. The first run downloads the model; later runs start instantly.
- Drag the comparison handle to inspect the result, then click Download PNG.
Specification
| Model | Real-ESRGAN x2plus / x4plus (ONNX) |
|---|---|
| Runtime | ONNX Runtime Web 1.27, WebGPU with WASM fallback |
| Input formats | PNG, JPG/JPEG, WEBP |
| Output | PNG, RGB, no watermark |
| Scales | 2× and 4× |
| Tiling | 256 px core, 32 px overlap |
| Upload | None — processing is local |
| Cost / limits | Free, unlimited, no account |
Frequently asked questions
Is this really free, and is there a watermark?
Yes, free, and no watermark. The tool does not use paid API calls — the model runs on your own hardware, so there is no per-image cost to pass on.
Are my images uploaded anywhere?
No. The image is read with createImageBitmap, processed in a Web Worker and encoded back to PNG inside the tab. No request carries image data. You can verify it in your browser's Network panel, or disconnect after the model has been cached and keep working offline.
Why is the first run slow?
The first run downloads about 69 MB of model weights and compiles the inference graph. That happens once per scale, then the weights are cached by the browser and later runs skip straight to inference.
Do I need WebGPU?
No, but it is much faster. Chrome, Edge and recent Safari support WebGPU. Without it the page uses a multi-threaded WASM build on the CPU, which produces the same output but takes several seconds per tile.
What is the maximum image size?
There is no hard limit — the constraint is your device memory. A 4× run on a 12-megapixel photo has to hold a 192-megapixel result in RAM, which most laptops will refuse. For very large sources, 2× is the realistic option.
Does it keep transparency?
Not yet. The model works on RGB, so the alpha channel is dropped and the output is opaque.
Will it fix a photo that is out of focus or heavily blurred?
Partly. Real-ESRGAN was trained on resolution loss and compression artefacts, not on motion blur or a missed focus. It sharpens and rebuilds texture; it cannot recover detail that was never captured.
How does it compare to Topaz Photo AI or Remini?
Those are paid, dedicated products with several specialised models and face-restoration pipelines. This is a single general-purpose model you can use immediately, for free, without installing or uploading anything. For casual enlargements the gap is small; for professional retouching work, it is not trying to replace them.