Files stay on your device. Enter send · Esc close drawers
RealESRGAN 2× — sharper than plain resize
AI image upscaling reconstructs detail that bilinear or bicubic resamplers can't recover. AI Upscale 2× runs the photo-tuned Real-ESRGAN model at 2× magnification, generating an output with sharper edges, cleaner text rendering, and reconstructed texture that matches the model's training domain. The model weights ship as a quantised ONNX file (~30 MB) and run in WebGPU where supported, falling back to WebAssembly otherwise — both happen entirely in your browser, so the source image is never uploaded. The preset outputs Photo upscale ~2× input.
| Processing | 100% in your browser — no upload |
|---|---|
| Cost | Free, unlimited |
This preset is a deterministic recipe — every step has fixed parameters so the result is byte-identical across runs of the same input.
AI Upscale 2× runs the Real-ESRGAN photo-2x model locally in your browser. The weights — about 30 MB quantised to fp16 — are fetched on first use, cached by the service worker, and re-used on subsequent runs without a network round-trip. WebGPU is the fast path: on a 2023+ Mac, Chrome PC, or Pixel-class phone the 2× upscale finishes in 4-8 seconds for a 1024×1024 input. Older hardware falls back to WebAssembly with multi-threaded SIMD; the result is identical, the wall-clock is 4-6× longer.
Because the model runs on your device, the source image never leaves the browser. You can verify this in the Network tab of devtools — only the static-asset and weights requests appear.
The photo-tuned variant of Real-ESRGAN. The photo model is trained on the corresponding content domain (natural photographs vs anime/illustration) and produces better edge reconstruction on that domain than the generic variant. Switch via the model dropdown if your source doesn't match.
On a recent laptop with WebGPU enabled, a 1024×1024 input upscaled 2× completes in 4-8 seconds. WebAssembly fallback (older browsers, no WebGPU) takes 4-6× longer. The preset shows a progress indicator while the model runs.
Yes — that's the point. 2× of a 1080p input is 3840×2160 px. Use upscaling when you're preparing for print, large-format display, or when you need to crop in further on a fixed-size source.
No. This preset runs entirely in your browser using WebAssembly + canvas APIs — no server round-trip. You can verify this in your browser's network panel: only static asset requests, no image upload. The file never leaves your device.