Does this upscaler upload my images to a server?

No. The Swin2SR AI model runs entirely in your browser using WebAssembly. Your images stay on your device at all times — nothing is transmitted to any server.

What is the maximum input image size?

Input images must be 2000×2000 pixels or smaller. For best performance and speed, we recommend images under 1000×1000 pixels. Larger inputs take significantly longer to process.

How long does AI upscaling take?

2× upscaling of a 500×500 px image takes roughly 10–30 seconds on a modern laptop. 4× upscaling can take 30–120 seconds. Processing time depends on your device's CPU speed — no GPU acceleration is available in the browser.

What is the difference between 2× and 4× upscaling?

2× doubles each dimension (a 500×500 image becomes 1000×1000), while 4× quadruples them (500×500 becomes 2000×2000). Both use different Swin2SR model checkpoints. 4× takes longer and produces a larger file. Choose 2× for most use cases; 4× is ideal for preparing low-resolution assets for large-format print.

AI Image Upscaler — 2× & 4× Super-Resolution

Enlarge images without the blurriness of traditional bicubic interpolation. This tool uses Swin2SR, a Swin Transformer-based super-resolution model trained on high-quality image pairs. It reconstructs fine details, sharp edges, and textures at 2× or 4× the original resolution.

Ideal for restoring old photographs, preparing low-resolution product images for print, upscaling screenshots for high-DPI displays, and rescuing assets created at too-small a resolution.

The AI model (~12 MB) is downloaded once and cached locally. Your image never leaves your device. Max input: 2000×2000 px.

Upscale factor

2× upscaling typically completes in 10–30 seconds.

How AI super-resolution works

Classical upscaling (bicubic, Lanczos) interpolates pixel values mathematically. The result looks smooth but lacks the detail that was never in the original. AI super-resolution instead uses a neural network trained on millions of high-resolution/low-resolution image pairs to learn what plausible high-resolution detail should look like.

Swin2SR uses a Swin Transformer architecture — the same class of model used in state-of-the-art image recognition — adapted for pixel-level generation. It processes overlapping patches of the image in parallel, producing a coherent high-resolution result even at 4× scale.

The model (~12 MB) is downloaded once and cached in your browser via the Cache API. All inference runs on your CPU through a WebAssembly-compiled ONNX runtime. No GPU is required, though processing time varies with CPU speed.

Frequently Asked Questions

Does this upscaler upload my images to a server?: No. The Swin2SR AI model runs entirely in your browser using WebAssembly. Your images stay on your device at all times — nothing is transmitted to any server.
What is the maximum input image size?: Input images must be 2000×2000 pixels or smaller. For best performance and speed, we recommend images under 1000×1000 pixels. Larger inputs take significantly longer to process.
How long does AI upscaling take?: 2× upscaling of a 500×500 px image takes roughly 10–30 seconds on a modern laptop. 4× upscaling can take 30–120 seconds. Processing time depends on your device's CPU speed — no GPU acceleration is available in the browser.
What is the difference between 2× and 4× upscaling?: 2× doubles each dimension (a 500×500 image becomes 1000×1000), while 4× quadruples them (500×500 becomes 2000×2000). Both use different Swin2SR model checkpoints. 4× takes longer and produces a larger file. Choose 2× for most use cases; 4× is ideal for preparing low-resolution assets for large-format print.