IT Image Tools

AI Image Upscaler — Upscale Images 2× or 4× in Your Browser

Enlarge photos 2× or 4× with Swin2SR AI super-resolution running locally via WebGPU or WebAssembly. Before/after slider, PNG download, Lanczos fallback — no uploads.

🔒 Runs entirely in your browser — nothing is uploaded

Drop an image here or click to browse

Your image is processed on this device and never uploaded. For AI mode the Swin2SR model weights are downloaded once from the Hugging Face hub and the ONNX runtime from the jsDelivr CDN; both are cached by your browser.

Advertisement

AI super-resolution without uploading your photos

Classic resizing stretches existing pixels, so an enlarged photo looks soft and blocky. An AI image upscaler instead uses a neural network trained on millions of image pairs to predict how sharp edges and fine textures should look at the higher resolution. This tool runs the open Swin2SR super-resolution models directly in your browser with Transformers.js and ONNX Runtime. The 2× mode uses the classical Swin2SR model, which is faithful to clean images such as renders, screenshots and good photos. The 4× mode uses the real-world Swin2SR model trained with BSRGAN-style degradations, which copes better with noise and JPEG compression in everyday pictures. Your image is processed on your own device and is never sent to a server.

Model download, speed and memory

On first use the model weights are downloaded once from the Hugging Face hub and cached by the browser: about 22 MB for the compact 8-bit version used on the WebAssembly (CPU) backend, or about 54 MB for the full-precision version used on WebGPU. The ONNX runtime itself comes from the jsDelivr CDN. If your browser supports WebGPU, the graphics card does the work and a typical photo is ready in seconds; otherwise the CPU is used, which is slower but works everywhere. To keep memory use low, the image is split into small overlapping tiles that are upscaled one after another and stitched back together without visible seams; a progress bar shows each tile. Transparent PNGs keep their alpha channel, which is enlarged with Lanczos filtering.

Fast fallback and fair comparison

If the model cannot be downloaded or your device runs out of memory, the tool automatically switches to Lanczos resampling, a high-quality classic filter that needs no download and is far better than the browser's default stretching. You can also pick it yourself for a quick result. When the job finishes, drag the before/after slider to compare the original, stretched to the same size, with the upscaled result, and switch to actual pixels to inspect details. The download is a lossless PNG at the new resolution.

How to use

  1. Add an imageDrop a JPG, PNG or WebP image or click to browse.
  2. Choose scale and methodPick 2× or 4× and AI super-resolution or fast Lanczos resampling.
  3. UpscaleWait for the model download (first time only) and the tile-by-tile progress.
  4. Compare and downloadDrag the before/after slider, then download the PNG.

Frequently asked questions

Is my image uploaded to an AI server?
No. The AI model runs inside your browser with WebGPU or WebAssembly. Only the model weights (about 22 MB, or about 54 MB on WebGPU) are downloaded once from the Hugging Face hub, and the ONNX runtime from the jsDelivr CDN; your image never leaves your device.
How large can the image be?
The result may be up to 4096×4096 pixels (about 16.7 megapixels) and 8192 pixels on the longest side, which covers 4× upscaling of images up to 1024×1024. Larger images are processed in tiles, but the final canvas must still fit in browser memory.
Why is it slow on my computer?
Super-resolution is heavy work. With WebGPU a typical image takes seconds; on the WebAssembly (CPU) fallback it can take a minute or more, especially at 4×. Progress is shown per tile and you can cancel at any time.
What if the AI model does not load?
The tool automatically falls back to high-quality Lanczos resampling, which needs no download. You can also choose the fast Lanczos method yourself.
Will upscaling add real detail?
The AI reconstructs plausible edges and textures learned from training photos, so results look much sharper than simple resizing. It cannot recover information that was never captured, such as unreadable text.
Advertisement