Enlarging an image is not a matter of making the pixels bigger. It is a matter of inventing pixels that were never captured. How convincingly software invents them is the entire difference between a blurry mess and a usable enlargement.
Why ordinary resizing goes soft
Classic resampling — bicubic, Lanczos, whatever your editor calls it — works by averaging. To place a new pixel between two existing ones, it takes a weighted blend of the neighbours. That is mathematically sound and completely uninformed: it has no idea whether it is looking at an eyelash, a brick edge or sky.
Averaging preserves smooth areas well and destroys edges, because an edge is precisely where neighbouring pixels disagree. Blend them and the edge softens. Do it at 4× and every fine detail turns to mush.
What AI upscaling does instead
Models such as Real-ESRGAN are trained on millions of pairs: a high-resolution image and a deliberately degraded version of it. The model learns what kinds of detail typically produce a given blurry patch, so instead of averaging it reconstructs plausible detail — hair as strands, brick as texture, text as strokes.
"Plausible" is doing real work in that sentence. The model is not recovering the original detail, because that information is gone. It is generating detail consistent with its training. For photographs this is usually indistinguishable from the real thing. For forensic or evidential purposes it is not a reconstruction and should never be treated as one.
Choose a scale that matches the source
| Source | Sensible scale | Expect |
|---|---|---|
| Sharp, well-lit, 1000 px+ | 4× | Excellent, print-ready |
| Decent phone photo | 2–4× | Very good |
| Small web image, 300–600 px | 2× | Good; 4× starts inventing |
| Heavily compressed or blurry | 2× | Cleaner, still soft |
| Tiny thumbnail under 150 px | — | Little to work from |
More scale is not better. Asking for 4× from a poor source gives the model licence to invent aggressively, which is where you get waxy skin and smeared, dreamlike texture.
Fix problems before enlarging, not after
Upscaling amplifies whatever is already in the frame. JPEG blocking, sensor noise and haloing all get enlarged along with the subject, and a model may interpret compression artefacts as texture and cheerfully reconstruct them at four times the size.
If you have any choice at all, start from the largest, least-compressed original you can find. An untouched 800 px original beats a 1600 px version that has been saved, resized and re-saved three times.
Upscale an image
Runs locally on our servers with Real-ESRGAN. No third-party API, and your file is not sent anywhere else.
Open Image Upscaler Remove backgroundWhere upscaling earns its keep
- Product photos — marketplace listings often demand a minimum resolution an older catalogue image cannot meet.
- Print — screen images are typically 72–96 DPI; print wants around 300. A 4× enlargement is roughly the gap.
- Old family photographs — scans of small prints respond very well, particularly combined with restoration.
- Logos and graphics — though if a vector original exists anywhere, use it instead. Vectors scale perfectly and infinitely.
Frequently asked
Can upscaling recover a face that is too blurry to recognise?
No, and you should be cautious of anything claiming otherwise. The model will produce a sharp, plausible face, but it is generating features consistent with its training rather than recovering the real ones. It can look convincing and still be wrong.
Should I upscale before or after removing the background?
Remove the background first, then upscale. Cutout edges are where quality problems show most, and upscaling afterwards refines that edge rather than enlarging a rough one.
Why does the result look smooth and slightly artificial?
Usually too much scale from too little source, or a model tuned for illustration applied to a photograph. Drop to 2× and compare; the difference is often immediate.