Image Upscaler

Classic resizing, whether bilinear, bicubic or Lanczos, can only interpolate between existing pixels, so enlarged photos look soft and JPEG blocks get magnified. Super-resolution models are trained on millions of pairs of low- and high-resolution images and learn what sharp edges, hair, foliage and text usually look like. They then predict that missing detail for your image. The result is not the true original detail, but it is usually far cleaner and sharper than interpolation.

Two models are available. Sharp uses Real-ESRGAN general x4v3 (BSD-3-Clause, about 5 MB), a compact convolutional network trained to remove blur, noise and compression artifacts while it upscales. It runs on any device. Faithful uses Swin2SR (Apache-2.0, about 54 MB), a transformer model that stays closer to the original pixels and invents less texture. The ×2 version is the classical model and the ×4 version is trained on real-world degradations. Swin2SR is much heavier, so it is only offered when your browser can use the graphics card through WebGPU. If the GPU can't run it, the tool falls back to Real-ESRGAN and tells you.

Large images are processed in tiles: 128 × 128 pixel blocks on the CPU and up to 192 × 192 on WebGPU, overlapping by 16 pixels and cross-faded so no seams appear. The progress bar counts tiles, so you can see how long is left. On a CPU, each tile takes roughly one to three seconds, so inputs are reduced to about 0.6 megapixels there. With WebGPU the only limit is the output size, which is capped at 16 megapixels and 8,192 pixels per side so every browser can still display and save it. For example, a 2,000 × 1,500 photo can be doubled but not quadrupled.

Transparent PNGs keep their transparency, because the alpha channel is enlarged separately with smooth resampling. Save as PNG for lossless output, or as JPG or WebP for much smaller files; a 16-megapixel PNG can easily exceed 30 MB, while a high-quality JPG of the same image is usually a fraction of that.