How to enlarge a photo without uploading it
A small photo blown up by plain resampling turns soft, because interpolation estimates each new pixel from its neighbours and never adds detail. The upscaler on this site runs a super-resolution network called Swin2SR inside your browser instead — this guide covers which engine to pick, what the fidelity slider does, and where the limits are.
How to enlarge a photo in four steps
- Drop in the image you want to enlarge. It is read into a canvas in the page and processed in memory; the application has no upload endpoint.
- Choose Enhance or Resize and set the scale. Enhance runs the Swin2SR network at 2x or 4x; Resize resamples with the browser built-in resampler and invents nothing. The model is fixed at 4x natively, so the 2x option runs the network and then resamples down.
- Tune fidelity and sharpening, then run it. Fidelity blends the AI estimate toward pixel-exact interpolation; sharpening finishes the result. The first run downloads the model once, and later runs reuse the cached copy.
- Compare and export. The before and after slider shows what the network actually added, and the result exports as PNG, JPEG or WebP.
Why the first run is slower. The model file downloads once on the first run and then runs entirely on your hardware — WebGPU on the graphics processor where available, WebAssembly on the CPU otherwise. Once cached, the same page upscales with no model download at all.
Which engine for which image
| Image type | Suggested engine | Why |
|---|---|---|
| Photographs | Enhance | Estimating texture is desirable; noise and grain benefit from it. |
| Grainy or compressed shots | Enhance | The network is trained to clean up exactly this kind of degradation. |
| Screenshots and UI | Resize | Interpolation keeps every edge where it belongs. |
| Documents and scanned text | Resize | A network may reshape letterforms, which changes meaning. |
| Logos, line art, diagrams | Resize | Crisp geometry is reproduced exactly, not approximated. |
| Barcodes and QR codes | Resize | Module boundaries must stay pixel-sharp to remain scannable. |
| Mixed photo and text | Enhance, low fidelity | Some neural detail without disturbing the type. |
What fidelity is for
A super-resolution model cannot recover detail that was never captured. Once pixels are gone they are gone, and what the network produces is an estimate built from patterns it learned during training. On a photograph that estimate is usually convincing; on a document, a logo or a face it can be quietly wrong.
The fidelity slider makes that trade explicit. At 100% AI you get the network's full output. Drag it toward 0% and the result blends back toward plain interpolation, which changes no detail it was not given. Somewhere between the two is usually the right answer for the image in front of you.
Limits worth knowing
- Output is capped at roughly 8 million pixels. About 2800 by 2800: a 4x upscale of a 700 by 700 source, or a 2x upscale of a 1400 by 1400 one. Larger images tile automatically with surrounding context so seams do not show.
- Invented detail is still invented. The network emphasises edges beyond the original rather than merely reproducing them, which is why the fidelity slider matters — and why text goes through Resize.
- Privacy is verifiable, not promised. Open the network tab before you start: every entry before the run is a download of library or model files, and during the run the list stays still. No pixel data is transmitted at any point.
- Export is PNG, JPEG or WebP. Pick PNG for exactness, JPEG or WebP when the file needs to travel.
Frequently asked questions
How do I enlarge a photo without uploading it anywhere?
Drop the image into the upscaler, pick Enhance or Resize, set 2x or 4x, and run it. The Swin2SR network executes inside your browser through WebGPU or WebAssembly, so the pixels never leave your device.
The only downloads are the library, the model weights and the page-view analytics.
Should I use Enhance or Resize for my image?
Use Enhance for photographs and grainy shots, where estimating texture is desirable. Use Resize for screenshots, documents, logos, line art, barcodes and anything with text, because interpolation changes no detail it was not given while the network can quietly reshape a glyph.
What does the fidelity slider actually change?
It blends between the network output and plain interpolation. At 100 percent AI you get the full estimate; dragged toward 0 percent the result moves back toward resampling that alters nothing.
Somewhere between the two is usually right, and the before and after slider is there so you can judge rather than trust.
How large can the enlarged photo be?
Output is capped at roughly 8 million pixels, about 2800 by 2800. That is a 4x upscale of a 700 by 700 source or a 2x upscale of a 1400 by 1400 one.
Larger images are tiled automatically with surrounding context so the seams do not show, and a single very large upload is refused with an explanation.
Can upscaling bring back detail the photo never had?
No, and no upscaler can. The model estimates plausible detail from what remains, guided by patterns learned in training.
On photographs the estimate is usually convincing. On faces, text and logos it can be subtly wrong, because it is inventing rather than recalling.