Resampling: making up pixels
An image is a grid of measurements. Resizing produces a different grid, and the new sample points almost never line up with the old ones — a pixel in the output may sit between four pixels in the input. The algorithm has to decide what colour belongs there. That decision is called resampling.
Nearest neighbour takes the value of the closest original pixel. It is fast, preserves hard edges exactly, and produces visibly blocky results on photographs. It is the right choice for pixel art and for anything where you want to preserve exact pixel values, and wrong for essentially everything else.
Bilinear interpolation averages the four surrounding pixels. Smooth, fast, and slightly soft.
Bicubic considers sixteen surrounding pixels with a weighting curve that slightly overshoots at edges, which reads as sharpness. It is the default in most editing software and a good general answer.
Lanczos uses a windowed sinc function over a larger neighbourhood and generally produces the sharpest result for photographic downscaling. Its overshoot can produce faint ringing next to very high-contrast edges, which is occasionally visible on text.
For most photographic work the practical ranking is Lanczos or bicubic for photographs, nearest neighbour for pixel art, and bilinear when speed matters more than quality.
Why downscaling is safe
Making an image smaller discards information, which sounds bad and is in fact the easy direction. Every output pixel is derived from real measured data — several input pixels averaged together — so nothing has to be invented.
Downscaling can even improve apparent quality. Averaging groups of pixels reduces noise, which is why a high-ISO photo often looks cleaner at half size. It is also why a slightly soft photo can look acceptably sharp when displayed small.
The one genuine pitfall is aliasing. Reducing dimensions dramatically in a single step can undersample fine repeating detail — a striped shirt, a brick wall, a fine grid — producing moiré patterns that were not in the original. Good resizing filters low-pass the image before sampling to prevent this. If a tool produces moiré on downscale, resizing in two stages usually avoids it.
Why upscaling is not
Enlarging is fundamentally different: the information you need was never captured. A 500-pixel-wide photo does not contain 1000 pixels of detail that a clever algorithm can extract. It contains 500, and the other 500 have to be fabricated.
Classical interpolation fabricates conservatively, by smoothly blending neighbours. The result is never wrong, exactly, but it is soft — an enlarged image with no more real detail than the original, spread over more pixels. Sharpening afterwards increases local contrast at edges, which reads as sharper without adding detail, and overdone produces halos.
This is why the honest answer to 'can you enhance this?' is usually no. The information is not there.
What AI upscalers really do
Machine-learning upscalers work differently, and the distinction matters. Rather than interpolating, they have been trained on millions of image pairs and have learned what kinds of detail tend to accompany what kinds of low-resolution patterns. Given a blurry region that looks like it was probably brickwork, the model generates plausible brickwork.
On many images the result is genuinely impressive and far better than interpolation. On others it is confidently wrong, and this is the important caveat: the model is generating detail that is statistically plausible, not detail that was there. Faces gain features that are not the person's. Text becomes crisp, legible, and says something slightly different. Textures gain a characteristic over-regular quality.
The practical rule: AI upscaling is excellent for making an image look good, and inappropriate where the image is evidence. Never use it on anything where the detail itself is the point — a document, a licence plate, a medical image, a screenshot being cited. Anything it adds is invention.
It is also worth knowing that these models are computationally heavy. Running one on a large image in a browser tab can lock the page for a long time, which is why our own upscaler uses high-quality Lanczos resampling rather than a generative model: it will not invent detail, and it will not freeze your tab.
Practical rules
The decisions that matter, in order:
- Always resize from the largest original you have, never from an already-resized copy.
- Downscale freely — it is safe and often improves apparent quality.
- Treat 2× as the sensible ceiling for upscaling anything that must look natural.
- Resize before compressing, not after: pixel count drives file size far more than the quality setting does.
- Preserve the aspect ratio unless you have a specific reason not to; stretched photos are immediately obvious.
- Sharpen after downscaling, not before — downscaling softens, and sharpening first amplifies noise you are about to average anyway.
- Never upscale anything where the fine detail is evidence.