Home › Guides › How to upscale an image
How to Upscale an Image Without Losing Quality
Why enlarging normally ruins a photo, what AI upscaling really does, and how to get the best possible result from a small original.
Why enlarging normally ruins a picture
A digital image is a grid of coloured squares. A picture 500 pixels wide holds exactly 500 columns of information and not one more. When you drag it larger in an ordinary editor, the software has to invent everything in between, and traditional methods do that by averaging neighbouring pixels. The result is mathematically sensible and visually disappointing: edges soften, fine texture vanishes, and small text turns into a grey smudge.
This is why an enlarged photograph so often looks wrong even when nothing is obviously broken. The detail is not blurred. It is simply absent.
What AI upscaling does differently
A super-resolution model is trained on enormous numbers of image pairs: a sharp picture, and a shrunken copy of the same picture. Over millions of examples it learns the relationship between the two, and therefore what a small pattern of pixels usually looks like when it is larger.
Given a new small image, it does not average anything. It predicts the missing detail based on everything it has learned about how edges, skin, fabric, foliage, metal and lettering behave. That is why the output has crisp edges and believable texture instead of a smooth blur.
The word "predicts" carries an important limitation. The model produces an informed reconstruction, not recovered data. It is very good at making a result that looks right. It cannot tell you what was genuinely there.
Choosing 2x, 3x or 4x
| Setting | Best for | Trade-off |
|---|---|---|
| 2x | Mild enlargement, most natural look | Fastest, smallest file |
| 3x | When 2x is not quite enough | Middle ground |
| 4x | Small originals that must become large, such as print | Slowest, largest file |
A common mistake is always reaching for the maximum. If you only need a moderate lift, 2x usually looks more natural and finishes considerably sooner.
Five things that improve the result
1. Start from the best original you have
Upscaling amplifies whatever is in the source, faults included. A heavily compressed JPEG carries blocky artefacts, and the model will faithfully enlarge those artefacts too. If you have an earlier, cleaner copy, use it even if it is smaller.
2. Crop first, then enlarge
If you only need part of the picture, crop before upscaling. The model then spends its effort on the part you actually want, and the job finishes faster.
3. Never upscale the output again
Running a result back through compounds the model's guesses. Edges begin to look plastic and textures turn repetitive. If you need more size, go back to the original and choose a larger factor once.
4. Match the format to the job
PNG preserves every pixel exactly and suits logos, screenshots and flat colour. JPG and WEBP give far smaller files and suit photographs. See our format comparison for a fuller explanation.
5. Be realistic about faces
General-purpose models rebuild facial detail plausibly but not accurately. For a very small face the result can look convincing while differing from the real person. Treat it as an illustration, never as evidence.
What upscaling cannot fix
- Motion blur and focus errors. That information was never captured.
- Severe compression damage. Blocks and colour banding get enlarged too.
- Missing content. Text too small to read will not become readable.
- Heavy noise. Grain may be sharpened along with everything else.
A sensible workflow
- Find the largest, cleanest original you have.
- Crop to the area you need.
- Choose the smallest enlargement that meets your requirement.
- Upscale once, then compare against the original at full size.
- Save as PNG for graphics, JPG or WEBP for photographs.
Ready to try? Open the upscaler and start with one image.