Double a picture’s size

At 2× each new pixel sits exactly halfway between two recorded ones. The kernel has close, relevant neighbours on every side and very little distance to bridge, so the error it makes is small and lands where the eye is least sensitive. Push to 4× and three quarters of every neighbourhood is itself invented; push to 8× and the arithmetic is mostly interpolating its own earlier answers.

This is why a print shop that tells you “we can go about double” is not being unhelpful. They are describing the point at which an enlargement stops being a reasonable presentation of the picture and starts being an interpretation of it.

Lanczos 3 is the default and is right for anything photographic. If the source is a screenshot, a logo or a diagram — flat colour with hard boundaries — switch to edge-guided, which measures which way each boundary runs before it interpolates and keeps a diagonal a line rather than a staircase. If it is genuine pixel art, use nearest: every soft pixel there is a lie about what the artist drew.

Keep the 1:1 strip in view while you decide. Its first tile repeats the pixels you had at the same size, which is the baseline every one of these methods is trying to beat, and the difference at 2× is usually obvious within one loupe position.

• Four times over →The usual answer when a picture is close but not close enough. Opens at 4×.
• Eight times over →The factor most likely to disappoint, offered anyway with the reasons attached.
• A rough small photo →Opens with the clean-up pass on, because a 600-pixel JPEG carries more than a picture.
• Where a trained model beats arithmetic, and where it doesn't →The case for the classical path, made fairly, including the cases it loses.