What this opens on
The PNG panel at 2× with edge-guided chosen and PNG out. When the picture has transparency, the stage paints a checkerboard behind it.
| Method | Good for | Transparency |
|---|---|---|
| Edge-guided | logos, icons, line art | premultiplied, clean edges |
| Lanczos 3 | photographs with a cut-out | premultiplied, clean edges |
| Mitchell | soft art, no rims | premultiplied, clean edges |
| Nearest | pixel art at a whole factor | each pixel repeated, alpha too |
Where this falls down
- A blurry logo stays blurry.Edge-guided keeps edges as sharp as the source has them. A vector original is always the better source for a logo.
- PNG files grow fast.Four times the pixels at 2× is roughly four times the file. WebP is the smaller choice where this browser writes it.
Questions people ask
- Why do enlarged PNGs often get a dark fringe?
- Because the resampler averaged the colour of fully transparent pixels — usually black — into the edge. Every method here multiplies colour by alpha before filtering and divides after, so the edge takes its colour only from the visible pixels.
- Which method is best for a logo?
- Edge-guided, which this page opens on: it tests which way each edge runs and interpolates along it, so lines stay lines. For pixel art at a whole factor, Nearest keeps every pixel square.
- What happens to transparency with an AI model?
- The models see three channels only. The enhanced result takes its transparency from the arithmetic pass, so the shape of the see-through area stays exactly where it was and the model only changes colour inside it.
- PNG or WebP?
- Both keep transparency. PNG is lossless and opens everywhere; WebP is smaller and offered only where this browser can really write it.
Better fit?
- Enlarge without losing qualityMitchell, no clean-up, no sharpening, halo guard on — and the overshoot measured on the 1:1 strip.
- Double the sizeThe factor with the least to apologise for. Opens at 2× on a sharp source.
- Enlarge a JPGSoften the JPEG block steps first, then enlarge; JPG out with the real file size measured.
- AI upscalingTwo trained models, downloaded once on request and run here. Invented detail, named as such.


