Blocky and small are not the same fault
Nearly every page that ranks for this treats pixelation and low resolution as one problem with one fix, and offers an upscaler for both. They are not one problem. An image can be blocky while holding plenty of pixels, and it can be perfectly smooth while holding almost none, and the two states are produced by different things and respond to different things.
Sorting yours into the right category first is worth more than any choice of tool, because the wrong category produces the wrong expectation. People run a genuinely small image through a deblocking tool and conclude that nothing works, or run a compression damaged one through an upscaler and get a larger copy of the same blocks. This page is about telling them apart, and about which parts of each are honestly recoverable.
Look at the edges, because each fault has a signature
Open the image at full size rather than at thumbnail size, and find a boundary where something light meets something dark. A face against a background works. So does text, so does a horizon.
If you see squares, arranged in a regular grid, all the same size, with flat colour inside each one, that is compression. The grid is the giveaway. Nothing photographed in the real world lines up that way, and the regularity exists because the file format divides the picture into fixed tiles and describes each tile separately. Asked to describe them cheaply, the tiles stop agreeing at their borders and you can see exactly where one ends.
If instead the edges are soft and stepped, with no grid and no flat squares, and the picture simply runs out of detail evenly everywhere, that is resolution. Nothing is damaged. There was never enough recorded in the first place.
A third case is worth naming because it is common and it looks like the first. An image that was small, then enlarged by something, then saved again carries both faults at once. It is the hardest of the three, since the enlargement invented detail and the second save then damaged the invention.
- Regular squares of flat colour, same size, lined up in a grid: compression damage.
- Even softness with no grid and no blocks, detail simply absent: not enough resolution.
- Coloured smears or halos hugging a hard edge, worst around text: compression again, and specifically the colour information, which is stored more cheaply than brightness.
- Blocks that are large and obviously deliberate, sitting over a face or a document: that is redaction rather than a fault, and nothing recovers it.
What a model can put back, and what it is inventing
This is the honest part, and the part the tool pages skip. A model trained on pairs of damaged and clean images learns what a damaged region usually looks like when it is clean. That is a genuinely powerful thing, and it is also, precisely, a guess about the most likely original rather than a recovery of the actual one.
The guess is excellent where the content is predictable. Skin, hair, fabric, foliage, sky. These have structure a model has seen an enormous amount of, so the reconstruction is usually both plausible and close to right, and the result reads as a better photograph rather than as a repaired one.
The guess is poor, and occasionally confidently wrong, where the content is arbitrary. Text, digits, a plate, a logo, a specific woven pattern, the exact geometry of a face that occupies only a handful of pixels. There is no most likely answer to fall back on here, so the model supplies something with the texture of detail but not the detail that was there. It will look sharper and it will not be the same information.
That distinction is the thing worth carrying away from this page. Use these tools freely on pictures of people and places. Do not use them to read something.
Deliberate pixelation does not come back, and it is worth knowing why
Pixelation is also applied on purpose, to hide a face, a document, a screen or a plate. People regularly ask whether an enhancer will undo it, and the answer is no in a way that deserves understanding rather than just accepting.
Redaction works by averaging. Each block is replaced with a single value standing in for everything that was inside it, and the individual values are thrown away. The information is not sitting behind the block waiting to be uncovered. It is gone, and there is nothing in the file for any tool to find.
What a model can do is generate something that fits the space: a face that could plausibly belong there, characters that look like characters. It will be confident and it will be fabricated. Treating that output as a recovery is the one genuinely harmful mistake in this area, which is why the answer here is a flat no rather than a qualified one.
Stop making it worse, which is most of the fix
Compression damage accumulates, and almost all of it is added after the picture was taken, by the ordinary business of moving a file around. A surprising share of pixelation problems are solved by going back one step rather than by processing forward.
- Find the original. A picture that reached you through a messaging app, a social platform, or a screenshot of a screenshot has been re-encoded at every step, and every step threw a little more away. The version still sitting on the camera or in the original email is frequently clean.
- Do not zoom in and screenshot. This is the most common way people create the problem for themselves. Enlarging on screen and capturing the result bakes the enlargement in permanently, and the blocks come along with it.
- Save once, at the end. Every re-save of a compressed format applies the damage again, and the loss is invisible on any single pass, which is exactly why it goes unnoticed across a dozen well meaning edits.
- Do not crop before enhancing. Cropping throws away the pixels the tool would have worked from, and it is a decision you can make afterwards on a copy for nothing.
- Run the enhancement on the largest and least handled version you have, and keep that version untouched afterwards so you can try again differently.
If it is small rather than blocky, that is a different job
An image with no grid and no blocks, which simply lacks detail, is not damaged and there is nothing to remove. What it needs is enlargement that generates plausible detail as it scales, and that is a different operation with different limits and a different set of honest expectations.
The two can be stacked, in one order only. Clean the compression first, enlarge afterwards. Doing it the other way round scales the blocks up along with everything else and hands the second step a worse starting point than it needed. The resolution guide covers what enlarging genuinely buys and where it stops, and the photo quality guide is the place to begin if you cannot yet tell which of several faults you are looking at.
What Clara does with each of these
Clara works on the image rather than on the object. Faces gone soft, files too small, pictures flat or faded, and black and white that wants colour. On a compression damaged photograph of a person it does well, because a face is exactly the kind of predictable content described above, and that is the same reason it is the wrong tool for recovering text or a plate.
It does not repair physical damage either, so a torn or creased print comes back sharper and still torn. If your picture has several things wrong with it at once and you are unsure which to attack first, the photo quality guide sorts them, and the enhancer app comparison covers what the different categories of tool are actually built to do.
More guides
Questions
Can a pixelated image really be fixed?
Partly, and it depends which fault you have. Compression blocks can be removed convincingly, because the tool is repairing damage done to information that was genuinely recorded. An image that simply lacks detail has nothing to repair, so anything added is generated rather than recovered. Both can look good. Only one of them is restoration.
How do I tell whether my image is pixelated or just low resolution?
View it at full size and look at a hard edge. Regular squares of flat colour in a grid mean compression. Even softness with no grid and no blocks means there were never enough pixels. The grid is the reliable signal, because nothing photographed naturally lines up that way.
Can AI unpixelate a censored face or a blurred plate?
No. Deliberate pixelation averages each block down to one value and discards what was inside it, so there is nothing left in the file to recover. A model can generate a plausible face or plausible characters, which will look convincing and will be invented, and it should never be treated as the original.
Why did my image get pixelated after I sent it?
Messaging apps and social platforms re-encode what you upload, and every re-encode applies compression again. The damage accumulates across steps, so a picture that has been forwarded, screenshotted and re-sent carries several rounds of it. Going back to the original file usually solves more than any tool will.
Should I upscale first or unpixelate first?
Clean the compression first, then enlarge. Enlarging a blocky image scales the blocks up along with the picture and hands the next step a worse input. The order matters more here than the choice of tool does.
Does zooming in cause pixelation?
Zooming alone does not, since you are only looking more closely at what is already there. Capturing that zoomed view does, because a screenshot bakes the enlargement into a new file permanently. Work from the original at its own size rather than from a capture of a magnified view.