Guides
How to Upscale Images Without Losing Detail
Why a plain resize looks soft, what an AI upscaler actually does, and when 8x is a bad idea.
Enlarging an image the old way interpolates between the pixels you already have. There is no new information, so the result is the same picture with softer edges. That is why a 2x resize in any image editor looks blurry no matter which algorithm you pick.
What an AI upscaler actually does
It reconstructs plausible texture and edges based on what it has learned images look like. The output genuinely has more detail than the input.
Match the model to the material
Photo upscalers preserve grain, pores and natural texture; applied to line art they smear the edges that make it line art. Anime and illustration upscalers keep lines hard and flat areas flat; applied to a portrait they produce plastic-looking skin. Using one model for both is the most common reason an upscale looks wrong.
How far you can push it
| Scale | Reliable on | Falls apart on |
|---|---|---|
| 2x | Almost anything, including text | — |
| 4x | Photographs, illustration, product shots | Very small faces, dense patterns |
| 8x | Smooth subjects, large simple shapes | Fine text, distant faces, repeating textures |
Upscale last, not first
Order matters more than people expect. Noise, compression blocks and colour casts all get reconstructed faithfully by an upscaler, which means it will happily enlarge the flaws along with the picture. Clean first, enlarge second. The one exception is a source so small that the cleanup has nothing to work with — under about 400px on the long edge, a 2x pass first can give the denoiser enough to grip.
Print is the other case where the sequence matters. Work out the pixels you actually need — inches times DPI, so an 8x10 at 300 DPI is 2400x3000 — and pick the smallest scale that clears it. Going further than the target does not improve the print and does add reconstruction the paper will show.
Judge it at full size
The artefacts that matter are invisible at preview size, which is why our AI image upscaler puts a before-and-after slider on the output and lets you inspect at 100%. If the problem is noise or fading rather than size, enhance a photo is the better starting point.
Tools this guide is about
- Upscale ImageSet the scale and the model here, then run it in the workspace and compare the result at full size. Separate models for photos and for line art, because one model cannot do both.
- AI Photo EnhancerFour fixes, each one you can turn off. Because a one-click enhancer that always does all four is how photos end up looking plastic.