Adding pixels and adding detail are different jobs
Almost every result for this is a tool, and almost every one of them promises a higher resolution image with no loss of quality. That sentence uses resolution to mean two different things, and the gap between them is where every disappointing result comes from.
One meaning is how many pixels the file contains. The other is how much of the subject is actually recorded. The first can be changed in an instant and is worth nothing on its own. The second cannot be changed at all, only estimated, and knowing which one you need decides whether any of these tools will help you.
Pixel count is the easy half and the meaningless half
Any image can be given more pixels. The software divides each existing pixel into several and works out intermediate values from the neighbours, so the file gets larger and the picture gets smoother. Nothing about the subject has been recovered, because nothing new was ever recorded.
This is why enlarging a small photo and then looking closely feels like being cheated. The blur did not go away, it was enlarged along with everything else, and it now occupies more of the screen than it did before.
What an upscaler adds is a plausible guess
A model trained on an enormous number of photographs does something different from smoothing. It proposes what the detail probably was, based on what detail of that kind usually looks like, and draws it in. The result genuinely does contain more information than the file it started from. The information is inferred rather than recovered.
That distinction sounds academic until you notice where the guess is reliable and where it is not, because the two are very far apart and they sit in the same picture.
- Texture is safe. Brick, fabric, foliage, hair, gravel and skin all behave statistically, so a plausible version of them is close enough to a true version that nobody could tell and nothing depends on the difference.
- Edges are usually safe. A roofline, a window frame or the boundary of a coat had a definite shape and the model has seen a great many of them.
- Faces are not safe. A face at low resolution constrains the answer far less than it appears to, and the model will produce a clear, confident face that belongs to nobody. This is the one part of the picture people are looking at.
- Text is not safe. Signs, number plates, handwriting on the back of a print. The model will render crisp letters that are the wrong letters, and crisp wrong is worse than blurred right.
Ask what the picture has to survive
The safe and unsafe cases only matter in proportion to what the image is for, and that is the question the tool pages never ask. An enlargement destined to be printed and hung is being looked at, not interrogated. Invented brickwork behind somebody is invisible and harmless, and the upscaled version is simply the better picture.
An image that has to establish something is a different object. Who is in the photograph, what the sign said, whose handwriting is on the back. There the invented detail is the answer, and a confident wrong answer is worse than an honest blur, because a blur announces its own uncertainty and a sharp face does not.
The same reasoning runs through what colourisation can and cannot know, which is the same problem in a different dimension.
Real resolution is almost always upstream
If the print, the slide or the negative still exists, that object holds detail your file does not, and going back to it is the only way to get resolution that is genuinely there rather than proposed. A careful recapture usually beats any amount of upscaling of a bad copy, and it costs an afternoon rather than a subscription.
This is the single most common miss. People upscale a low quality phone snap of a print that is sitting in a drawer in the next room. The sequence that works starts with making one honest copy at the best size the equipment truly resolves, and the restoration guide covers how to make that copy without baking decisions into it.
Where no better original exists, an upscaler is the right tool and there is nothing dishonest about using it. Just know that you are commissioning an illustration of the subject rather than recovering the subject.
How far to push it, and how to tell
There is no universal ceiling, because the limit is set by how much real detail the file already carries rather than by a multiplier the tool advertises. A sharp small image tolerates a lot of enlargement. A soft one runs out almost immediately, and pushing it further only makes the invention more elaborate.
The test that works is to look at the part that matters, at the size you will actually view it, and ask whether you believe it. Compare the result against the original at the same scale. If a face has gained detail that was not visible before, that detail was authored rather than found, and you can decide whether you mind.
Keep the original file either way. Every enlargement is a version, and a version should never be the only copy left.
More guides
Questions
Can you increase image resolution without losing quality?
You can increase pixel count without losing anything, because nothing is being discarded. What you cannot do is add real detail. Anything sharper than the original is inferred, which is fine for texture and unreliable for a face or a line of text.
Why does my upscaled photo still look blurry?
Because simple enlargement smooths between the pixels you already had rather than adding any. The softness was in the original and it has been enlarged with everything else. A model based upscaler will look sharper, though what it added is an estimate rather than a recovery.
Is it better to rescan a photo than to upscale it?
Almost always, if the print or negative still exists. The physical original holds detail no file of it has captured, so recapturing it gives you resolution that is really there. Upscaling is the right answer only when there is no better original to go back to.