Upscale images with AI, right in your browser.

Getting the best results from AI upscaling

Upscaling is a judgment tool, not a magic button. The gap between a crisp result and a mushy one comes mostly from the choices you make before you press run — which scale fits your source, whether the model has enough to work with, and what you do with the bigger file afterward. Here is how to get a clean result and keep your expectations honest along the way.

×2 or ×4: which to choose

×2 is the default, and usually the right call. It doubles each side for four times the pixels, runs on the smaller of the two models (around 8 MB), and finishes sooner — a good fit when the source is already decent and you just want a cleaner, larger copy. ×4 quadruples each side, sixteen times the pixels, and leans on a much larger model (around 55 MB) that takes longer to download the first time and longer to run. Save it for sources that are genuinely small or low-resolution. Not sure which is right? Run ×2, then toggle ×4 to compare — each scale is cached, so switching back and forth is instant and never re-processes an image you have already done.

What upscaling reconstructs — and what it can't invent

The model was trained on millions of low- and high-resolution image pairs, so it predicts the most likely detail rather than averaging pixels the way a plain resizer does. That means it genuinely rebuilds edges, texture, and some of the softness left behind by compression. What it cannot do is recover detail that was never captured. A face a dozen pixels wide has no features to restore, and the model may invent uncanny ones; small text can get subtly reshaped into letters that were never typed. This is reconstruction, not clairvoyance, so always start from the largest, cleanest version you have — never a tiny copy of a copy.

Best inputs, and the ones that fight back

Some sources are exactly what the model was built for; others actively work against it.

• Upscale well — clean photos that are merely a bit soft or small, illustrations and line art, logos and flat graphics, product shots, and AI-generated renders.
• Fight back — heavily-compressed JPEGs, where the model may sharpen the blocky artifacts instead of hiding them; screenshots of text and interfaces; and very noisy or grainy images.
• A quick tip — for a noisy or over-compressed source, ×2 often looks more natural than ×4, because it commits to less and exaggerates the flaws less.

Pairing upscale with resize and compress

More pixels means a heavier file, so upscaling is rarely the last step. Upscale first, then run the result through the compressor to bring the bytes back down — the same reconstructed detail at a fraction of the size, which matters most for anything headed to the web. Going the other direction, if all you need is a smaller image, don't upscale at all: resize with a good downscaling filter is the right tool, and far faster. A common flow is to upscale a small source up to a usable size, resize it to the exact dimensions you need, then compress for weight.

It runs in your browser — even the heavy model

The whole pipeline runs on your device. The model runs on WebGPU where your browser supports it and falls back to WebAssembly on the CPU otherwise, and even the larger ×4 model downloads once and is cached for every run after. Inference happens in a background worker, so a slow ×4 or a big tiled job never freezes the page — you can queue a whole batch and keep working while it churns. Nothing is uploaded; no server ever sees your images.

For the fuller picture of how neural super-resolution differs from a plain resize — with worked examples and a megapixel comparison — read How AI image upscaling works.