AI Photo Upscaling: Fix Blurry Photos in Seconds
What upscaling can and cannot recover, why it invents detail rather than finding it, and how to get a usable result from a small or soft original.
Upscaling is the most misunderstood tool in the set, because what it appears to do and what it actually does are different things. It appears to recover detail. It does not. It predicts detail that is consistent with what is there, and paints it in.
That distinction is not pedantry. It determines when the tool is brilliant and when it quietly lies to you.
What it actually does
A traditional resize is arithmetic. To double an image's width, the software has to invent a pixel between every existing pair, and it does so by averaging. The result is bigger and blurrier, because averaging is the opposite of detail.
An AI upscaler is a model that has seen millions of pairs of images — the same scene at low and high resolution — and has learned what tends to be present in the high-resolution version. Given a soft edge, it does not average; it predicts the sharp edge that most likely produced that softness, and draws it.
When the input contains enough signal, the prediction is close to the truth and the result is genuinely excellent. When it does not, the prediction is still confident, and the result is a well-rendered fiction.
The cases where it works very well
- Slightly soft photos. A photo that missed focus by a little, or was taken at a high ISO and denoised into mush. There is real structure under the softness for the model to sharpen.
- Small originals. An old 800-pixel-wide JPEG that you need at print size. Two to four times up is comfortable.
- Heavy crops. You cropped in hard on a good photo and ended up short of pixels. This is the most common everyday use and the one with the best hit rate.
- Scanned prints. Scanner softness is uniform and predictable, which is exactly what these models handle well.
The cases where it fails
- Genuinely blurred faces. If you cannot tell who it is, the model cannot either. It will produce a sharp, confident, incorrect face.
- Motion blur. The information was smeared at capture. Resolution was not the problem, so more resolution is not the answer.
- Text and numbers. Small text upscales into characters that look right at a glance and are wrong on inspection. Never upscale a document, a receipt, or a licence plate and treat the result as evidence.
- Repeating fine patterns. Fabric weave, brickwork, chain-link. Models regularise these, and the regularity is visibly artificial.
- Anything already upscaled. Running a second pass compounds the invention. One pass, from the best source you have.
Start from the best original
This is most of the battle, and it is not glamorous.
Go back to the original file. Not the version shared in a messaging app, which was recompressed. Not a screenshot of it. Not a download from a social platform, which resized it. Each of those steps threw away information permanently, and upscaling cannot get it back — it can only invent a replacement.
If you have a physical print, scan it rather than photographing it. A flatbed scan at 600 dpi gives the model far more to work with than a phone photo of the same print, and it avoids the perspective distortion and uneven lighting that a handheld shot introduces.
Check the result properly
Zoom to 100 percent and look at four places:
- Skin. Should have pores and fine lines. If it looks like wax or porcelain, the model has smoothed rather than sharpened.
- Hair. Individual strands, not brush strokes. Strokes mean the model was guessing.
- Eyes. The iris should have visible structure. Reflections should sit in the same place as the original.
- Edges against the background. Look for a faint bright line — a halo — which is the classic over-sharpening tell.
Then do the comparison that matters: open the original beside the result and ask whether anything has changed, not just whether the result is sharper. Sharper and different is a failure, even if it looks better.
Where it fits in a workflow
Upscaling is a separate step, done first. If you plan to also replace a background or remove an object, do the upscale before those, so the later tools have more detail to work with — and do each as its own operation rather than asking one prompt to do everything. Bundled instructions are how you end up unable to tell which step introduced the artifact.
For portraits specifically, upscaling a soft source before a headshot pass meaningfully improves the outcome, because the headshot step is then working from structure rather than from mush.
The honest summary
Upscaling is excellent at making a decent photo bigger and crisper. It is unreliable at rescuing a photo that never had the detail in the first place, and it is confidently wrong in exactly those cases. Use it freely on crops and soft shots. Use it carefully on faces. Do not use it on anything where being wrong matters.
Four of Kitana's tools run in the browser, and the full thirteen — including restoration for damaged prints — are in the iOS and Android apps.
Frequently asked questions
- Can AI upscaling recover a face that is completely blurred?
- No. It will produce a sharp face, but not that person's face. When the input has no facial detail, the model fills the gap from what faces generally look like, and the result is a plausible stranger. This is the single most important limit to understand: upscaling reconstructs, it does not retrieve.
- What is the difference between AI upscaling and just resizing?
- Traditional resizing interpolates — it averages neighbouring pixels to invent in-between ones, which makes an image larger and softer. AI upscaling predicts what detail should be there based on millions of examples, which makes it larger and sharper. The sharpness is generated, not recovered, which is why it can be both impressive and wrong.
- How much can I enlarge a photo before it falls apart?
- Two to four times the original dimensions is the reliable range for most photos. Beyond that the model is inventing more than it is reconstructing, and the failure modes get obvious: waxy skin, hair that turns into brush strokes, and text that becomes convincing gibberish.
- Does upscaling fix motion blur or camera shake?
- Usually not well. Upscaling addresses resolution, and motion blur is a different problem — the information was smeared across pixels at capture, not simply discarded. Some tools attempt deblurring too, but results on genuine motion blur are far less reliable than on soft or small images.
- Will upscaling make a screenshot look like an original?
- It will improve it, but screenshots carry compression artifacts on top of the resolution loss, and upscaling sharpens those artifacts along with everything else. Always go back to the source file if one exists. A screenshot of a photo is two generations of loss.
- Is it safe to upscale old family photos?
- It is safe, but be deliberate about it. Upscaling a scanned photograph of a relative will invent detail in their face that was never in the original. For a print you want to hang, that may be fine. For a historical record, keep the unmodified scan alongside it and label which is which.
Ready to put this into practice?
Create with Kitana using the tool that fits this guide.
Upscale a photoReady to try it yourself?
Download Kitana and create your first AI photo in under a minute.