FREE · LOCAL · NO UPLOAD

AI Image Upscaler

A roughly 4.9 MB ONNX model runs in a Web Worker. The model natively produces 4x; 2x is downsampled from the AI 4x result.

Your image never leaves your device
The image is processed only in this browser tab and is never uploaded.

To prevent OOM, files are limited to 30 MB, input and output to 36 megapixels, either side to 16,384 px, with working memory estimated before allocation.

GUIDE & TECHNICAL NOTES

How to use AI Image Upscaler

Use Real-ESRGAN general x4v3 for tiled 2x or 4x browser upscaling with a fast interpolation fallback.

  1. 01Choose an image and a 2x or 4x factor.
  2. 02Keep AI mode for verified tiled inference, or choose fast interpolation.
  3. 03Review the AI/fallback badge, dimensions, and preview before download.

Real on-device AI super-resolution

Real-ESRGAN general x4v3 performs overlapping 4x ONNX tile inference in the browser. WebGPU is preferred and WASM is the fallback. For 2x, each AI 4x tile is reduced before stitching.

Memory and truthfulness limits

AI output is capped at 16 megapixels, working memory is estimated, and Tensor resources are released tile by tile. A limit or model failure switches to existing fast interpolation. Super-resolution predicts texture and may create artifacts; it is not recovered factual evidence.

Frequently asked questions

Is this AI super-resolution?

Yes. The default is Real-ESRGAN general x4v3 ONNX. The result identifies AI and WebGPU/WASM, while a failure is clearly labeled as fast fallback.

Why can a large photo not use AI 4x?

A 4x image has 16 times as many pixels. The AI pipeline caps total output and uses tiles to prevent browser OOM.

More image tools