How to Enlarge an Image Without It Turning to Mush
Enlarging a photo means inventing pixels that were never captured. What AI upscaling can genuinely recover, what it cannot, and when to fix the source instead.
Updated August 11, 20265 min read
There is a fundamental limit here that no tool escapes: a 400 × 300 photo contains 120,000 pixels of information, and enlarging it to 1600 × 1200 requires 1.9 million. The other 1.8 million have to come from somewhere, and that somewhere is guesswork. What has changed is the quality of the guessing — a model trained on millions of photographs makes far better guesses than the arithmetic that came before it. Knowing where that guessing succeeds and where it invents things is the whole skill.
Traditional scaling versus learned scaling
Classic interpolation — bicubic, Lanczos and their relatives — fills each new pixel by averaging its neighbours. It is fast, predictable, and produces a soft result, because averaging is by definition a blur. Enlarge four times this way and everything looks like it is behind frosted glass.
A learned upscaler works differently. Having seen millions of examples of what a sharp photograph looks like next to its downscaled version, it predicts what the missing detail probably was. Given a blurry edge it produces a sharp edge, because in its training data blurry edges at that scale were sharp edges. The results on real photographs are dramatically better than interpolation.
What upscales well and what does not
| Content | Result | Why |
|---|---|---|
| Landscapes, nature | Excellent | Textures are statistical — plausible detail is indistinguishable from real |
| Portraits (moderate enlargement) | Good | Faces are heavily represented in training data |
| Portraits (large enlargement) | Uncanny | Skin becomes waxy; features drift from the real person |
| Product photos | Good | Clean edges and even lighting are the easy case |
| Text and logos | Poor | The model invents letterforms that were never there |
| Line art and diagrams | Variable | Better handled by re-exporting from the vector source |
| Heavily compressed images | Poor | It faithfully enlarges the compression artefacts |
That last row is the most common disappointment. Feeding a small, heavily compressed image to an upscaler produces a large image with large, crisp compression artefacts. The blocking becomes more visible, not less, because the model reads it as detail worth preserving.
A workflow that gets better results
Find the best available source
Before upscaling anything, check whether a larger original exists — in your camera roll, in the email it arrived in, on the original website. Ten seconds of looking beats any amount of reconstruction.
Clean up before enlarging
Reduce noise with the noise reduction tool first. Enlarging noise turns fine grain into large blotches, and the model will treat that grain as detail to reconstruct.
Crop before enlarging
If you only need part of the image, crop first. Upscaling the whole frame and then cropping wastes effort on pixels you discard.
Enlarge in one step, not several
Two rounds of 2× is worse than one round of 4×. Each pass treats the previous pass's inventions as ground truth and compounds them.
Sharpen last, lightly
A small amount from the sharpener adds bite. Too much creates halos around edges, which is the tell that gives away a heavily processed image.
AI detail reconstruction on your own device (devices that can't run the model fall back to standard scaling), up to 4×. The photo is never uploaded, and there is no queue or credit limit.
Upscale an imageHow far can you push it?
As a rough guide: 2× is safe on almost anything and usually looks like a straightforwardly better image. 4× is good on photographic content with clean edges and no visible compression. 8× and beyond is a stylistic choice — the output is more the model's work than yours, and it shows on close inspection. Our upscaler goes up to 4×, which is where the useful range ends anyway.
Match the target to the use. A web image displayed at 800px wide does not need more than 1600px even on a high-density screen. A print at A4 wants about 3,500px on the long edge at 300 DPI; 2,500px (about 210 DPI) is the floor for prints viewed at arm's length. Upscaling far past what the destination will show buys nothing except processing time and file size.
When not to upscale at all
- Logos and icons. Ask for the vector source. An SVG or EPS scales perfectly to any size, and someone almost always has one.
- Screenshots. Retake them at a higher display resolution or zoom level. That takes a minute and produces genuinely sharp pixels rather than predicted ones.
- Anything evidential. Upscaled detail is invented. It has no place in documentation, identification or anything that will be relied on as a record.
- Text documents. Re-export the PDF at a higher DPI instead. The vectors are still in the file and will render sharply at any size.
- Images that only need to look sharper. Sometimes the problem is contrast rather than resolution. Try the image enhancer before reaching for reconstruction.
The honest summary: upscaling is very good at making photographs bigger and acceptable, and it is not a substitute for capturing enough pixels in the first place. Treat it as a rescue operation for images you cannot recapture, not as a routine step in a workflow you control.
Frequently asked questions
Can you really enlarge an image without losing quality?
Not in the strict sense — the extra pixels are predicted rather than recovered, so no new real information appears. What a good upscaler does is produce plausible detail that reads as sharp, which on photographic content is often indistinguishable from the real thing at normal viewing distances.
How much can I upscale an image?
2× is safe on almost any photograph. 4× works well on clean, uncompressed sources. Beyond 8× the output is largely the model's invention and will not hold up to close inspection, particularly on faces and text. Our upscaler stops at 4×.
Why does my upscaled image look plastic?
Usually too large an enlargement factor, or the source was already compressed. Models smooth away what they read as noise, and on skin that produces a waxy look. Try a smaller factor, clean up noise before enlarging, and add only light sharpening afterwards.
Can upscaling recover text in a blurry photo?
It will produce letter-shaped marks, but they are predictions rather than the original characters, and they are frequently wrong. For text, run recognition on the best original you have instead of enlarging a degraded copy.