What enlarging can and cannot do
Be clear about this before you start, because the expectation is what leads to disappointment.
Every pixel in an enlarged image is a weighted average of pixels that were already in the original. There is no new information. If the detail was not recorded, it is not there, and arithmetic cannot conjure it.
What a good resampler does is avoid adding blur. That sounds modest and it is not: the difference between a properly enlarged photograph and one dragged into a bigger box is immediately visible, and it is the difference between usable and not.
Why the browser's own result is soft
Draw an image into a larger canvas and you get bilinear interpolation, which blends the four nearest pixels. Blending is blurring, so bilinear enlargement is soft by construction.
Lanczos looks at a much wider neighbourhood and weights it with a windowed sinc function. The important part is that the weights go negative either side of the peak: a pixel two places away pulls the result slightly away from itself, which sharpens the transition instead of smearing it.
That is why the two panels above are worth looking at side by side. Same crop, same size, one through the browser and one through Lanczos, at actual pixels. It is not a subtle difference on anything with an edge in it.
Pixel art is a completely different problem
Everything above is wrong for a sprite.
Pixel art is defined by its hard edges and its small, deliberate palette. Blending destroys both: the crisp blocks turn to mush and the palette fills with colours the artist never chose. Running a photographic resampler over a sprite is the quickest way to ruin one.
Nearest neighbour is the safe answer. Every pixel becomes a hard block, no blending at all, no new colours. It is exactly correct and it keeps the diagonal staircases.
Scale2x is better where it applies. It comes from arcade emulation, where people needed sprites drawn for a 1990s screen to hold up on a modern one. Each pixel becomes four, and where two adjacent neighbours agree while the opposite pair do not, the corner takes their colour, which rounds a staircase into a diagonal. Flat areas and straight edges are left completely alone, and it introduces no colour that was not already in the image.
The tool looks at the image when you open it, counts distinct colours and how many neighbouring pixels are identical, and suggests Scale2x when that pattern says sprite rather than photograph. It is a suggestion, not a decision, and you can override it.
The AI question, answered honestly
Upscalers that claim to add detail are not lying. They do add detail. The detail is invented.
A model trained on millions of photographs looks at your blurry patch, recognises it as probably being an eye or brickwork or grass, and draws in a plausible version. On a landscape or a texture the result can be genuinely excellent and nobody is harmed by the invention.
On anything where the answer matters, it is dangerous. A blurred number plate becomes a sharp, confident, possibly wrong number plate. A face becomes a sharp face that is subtly not the same person. A serial number becomes legible and incorrect. Researchers have demonstrated this repeatedly, and it looks exactly as convincing when it is wrong as when it is right.
So: never on evidence, documents, identification or anything a decision rests on. This tool does not do it, which is a deliberate choice rather than a limitation to apologise for.
Enlarge last, and only once
Do the colour work, the cropping and the corrections at the original size, then enlarge as the final step.
Enlarging first means every subsequent operation runs on four or nine times as many pixels for no benefit, and any sharpening you applied gets softened by the resample anyway. Enlarging twice in a row compounds the softness rather than sharing it.
And save as PNG rather than JPEG if there is any chance of further editing. An enlarged image is already working with less information than it looks like it has, and a lossy re-encode on top of that is the point where it starts to look processed.
Common questions
Can this add detail that was not in the photo?
No, and nothing here pretends otherwise. Every pixel in the result is a weighted average of pixels that were already there. What a good resampler gives you is the same information presented without the extra blur a naive enlargement adds. That is a real and visible improvement, and it is not the same as recovering detail.
What about the AI upscalers that claim to add detail?
They genuinely do add detail, and what that means matters: the detail is invented. A model trained on millions of photographs guesses what the missing pixels probably looked like and draws them in. For a landscape that usually looks great. For a face, a number plate, a serial number or anything where the answer matters, it is making something up that will look convincing and may be wrong. Never use one on evidence, documents or identification.
How far can I enlarge before it looks bad?
Doubling is comfortable and most people would not identify the result as enlarged. Between two and four times the softness starts to show, particularly on hair, foliage, text and fine texture, though it is fine for a background or something viewed at arm's length. Past four times each original pixel is covering more than sixteen output pixels and no amount of processing will fill them convincingly.
Why does my pixel art turn into a blurry mess?
Because the usual algorithms blend, and blending is exactly wrong for a sprite: the hard edges are the artwork. Use Nearest neighbour to keep every pixel as a crisp block, or Scale2x, which keeps edges hard while rounding off the diagonal staircases. The tool tries to spot pixel art when you open it and selects Scale2x for you.
What is Lanczos and why is it better than what my browser does?
The browser scales with bilinear, which blends at most four neighbouring pixels, and blending is blurring. Lanczos weights a much wider neighbourhood with a windowed sinc function that has negative lobes either side of the peak, and those negative lobes are what keep edges crisp. It is what ImageMagick and every serious image library use. The comparison on this page shows both on the same crop so you can see it rather than take it on trust.
Why does the preview show a small crop instead of the whole image?
Because enlarging can only be judged at full size. Shrinking a preview to fit the screen averages pixels together and hides precisely the softness you are trying to assess, so a scaled-down preview of an enlargement always looks better than the file you get. Both panels show actual pixels, and you can click either to look somewhere else.
Is my image uploaded?
No. It is decoded, enlarged and saved on your own device. There is no queue, no account and no size limit beyond what your machine can hold.