Enlarge a small image without sending it anywhere. Pick a scale, pick a kernel — Lanczos-3 for sharpness or bicubic for smoothness — add a touch of sharpening, and compare the result against the original with the before/after slider. The resampling is written into the page rather than delegated to the browser's own scaler, which is a hint the specification lets each browser answer differently, so the output is the same wherever you run it. Drop in a folder and the whole batch is upscaled and zipped. One honest caveat: interpolation smooths, it does not invent. A 100×100 photo enlarged to 800×800 is a larger file with the same information in it, and no amount of resampling changes that.
| Kernel | Character | Best for |
|---|---|---|
| Lanczos-3 | Sharpest, slight ringing | Photographs, detailed images |
| Bicubic (Catmull-Rom) | Smooth, no ringing | Graphics, screenshots, line art |
| Nearest neighbour | Blocky, no blending | Pixel art, seeing the difference |
No, and be wary of any tool that claims otherwise without a model. Interpolation redistributes the pixels that are already there. It makes the enlargement smooth rather than blocky, which is usually what people want, but the information content does not increase.
Lanczos-3 is sharper and is the right default for photographs and detailed images. Its trade-off is slight ringing — a faint light/dark halo — next to very hard edges, so for flat graphics, screenshots or line art bicubic often looks cleaner.
Two reasons. It shows you what no interpolation looks like, so the other two have something to be compared against, and it is genuinely the correct choice for pixel art, where smoothing destroys the whole point.
No. It is decoded, resampled and re-encoded in your browser, and nothing is sent to a server.
Up to roughly 16384 pixels per side and 40 megapixels in total, which are the browser's canvas limits rather than ours. Ask for more and the scale is capped with a note explaining why, because the alternative is a canvas that silently returns a blank image.
A separable Lanczos-3 pass touches every output pixel six times per axis, so an 8× enlargement of a 2-megapixel photo is a few hundred million multiplications. It runs to completion; it just is not instant.
Not here. Super-resolution models genuinely do reconstruct plausible detail, but they mean a model download of tens of megabytes before the first image, and the detail they add is invented rather than recovered. This tool does the honest, instant, deterministic thing.
Yes, and correctly. Colour is premultiplied by alpha before filtering, so soft PNG edges do not pick up the dark halo that a naive resampler produces. Export as PNG or WebP to keep it; JPEG has no alpha channel.
The resampling kernel is written out here rather than handed to the browser, and the reason is reproducibility. Canvas exposes only `imageSmoothingQuality: "high"`, which the specification defines as a hint: Chrome, Firefox and Safari each answer it differently, and Safari on some hardware ignores it altogether. An upscaler whose output depends on the visitor's browser is not a tool. So the image is resampled with an explicit separable two-pass convolution, offering Lanczos-3 for maximum sharpness, Catmull-Rom bicubic for a smoother result without ringing, and nearest-neighbour so you can see what no interpolation looks like. Two details in that pipeline are where most hand-written resamplers go wrong. The first is pixel centres: mapping output index i to source position i×ratio rather than (i+0.5)×ratio−0.5 shifts the whole image by half a pixel, which reads as a soft edge on one side and a hard one on the other. The second is alpha. Colour is premultiplied by alpha before filtering and divided back out afterwards, because a fully transparent pixel in a PNG still carries arbitrary RGB — usually black — and filtering it alongside its opaque neighbours drags that black into every soft edge as a dark halo. Kernel weights are normalised per output pixel so overall brightness is preserved, and samples that fall outside the image clamp to the edge rather than being dropped, which would pull the border inwards and leave a visible rim. Sharpening is an unsharp mask on the colour channels only, applied after resampling; sharpening alpha would carve notches out of anti-aliased edges. Finally, the scale is clamped before anything is allocated: browsers refuse a canvas beyond roughly 16384 pixels per side and a total area limit, and the failure mode is a silently blank image rather than an error, so the cap and its reason are shown instead. None of this invents detail. Interpolation redistributes the information already present; it cannot recover what was never captured.