Computed, not recorded
Interpolation fills between samples. It does not recover absent ones. The share of the output that can sit on a recorded sample falls with the square of the factor, and the lab prints it beside every enlargement:
| Factor | Output pixels per source pixel | Recorded | Computed |
|---|---|---|---|
| ×1.25 | 1.56 | 64.0 % | 36.0 % |
| ×1.5 | 2.25 | 44.4 % | 55.6 % |
| ×2 | 4 | 25.0 % | 75.0 % |
| ×3 | 9 | 11.1 % | 88.9 % |
| ×4 | 16 | 6.25 % | 93.8 % |
| ×8 | 64 | 1.56 % | 98.4 % |
That is why the buttons here never say “enhance” or “HD”. A larger file with the same information is still useful — a print vendor’s minimum, a layout that needs the pixels — and the receipt records the kernel that made it.
Limits
- Above about ×4 the result looks soft whichever kernel is used: the kernel can only blend what is there.
- Sharpening is meant for reductions. On an enlargement it exaggerates the kernel's own transitions rather than the picture.
- The 60-megapixel ceiling is per output, so a 24-megapixel photograph can be enlarged only about ×1.58.
Questions
Does upscaling add detail?
No. A classical kernel computes each new pixel from the recorded ones around it. The picture gets more pixels and smoother transitions between them, not information that was never captured. The detail meter on the Scale panel says what share of the output is recorded and what share is computed.
Which kernel should I use to enlarge?
Bicubic (Catmull-Rom) is the default: it passes through every source sample and stays crisp. Mitchell rings less beside hard edges at the cost of a little softness, which suits faces and skies. Lanczos is the sharpest and rings the most. Nearest is for pixel art at whole-number factors. The kernel strip shows them all on a patch of your own picture.
How far can a picture be enlarged here?
Up to ×16, or until the result reaches 60 megapixels or 16,384 px on a side — the size past which a browser canvas fails without saying so. The panel names the largest factor your picture can take before you press anything.
Why does the meter say 94 % at ×4?
At ×4 each source pixel becomes a 4 × 4 block of sixteen output pixels, and at most one of those sixteen can sit on the recorded sample. One in sixteen is 6.25 %, so about 94 % of the result is interpolated.
Is there an AI upscaler?
Not in this version. Model-based reconstruction predicts detail rather than recording it, so when it ships it will be a separate, labelled mode, never mixed into this one. Everything on this page is deterministic: the same file and settings give the same bytes.
Neighbouring measurements
Enlarging a sprite? Nearest at a whole number keeps every block square. Enlarging for a print? Check first whether only the DPI needs to change.
- Resize pixel artWhole-number nearest: every block square, no colour invented.
- Resize an image for printingA4, photo and frame sizes: retag, lower DPI, or short by N pixels.
- Image resamplingA percentage-led pixel resize with the interpolation cost stated before export.
- Why upscaling loses detailInterpolation fills between samples. It does not recover absent ones.
