Unblur Image: Fixes That Work by Blur Type

Because blur is not a layer sitting on top of a sharp picture. It is a record of something the sensor never resolved, and the amount recoverable depends entirely on how the detail was lost in the first place.

Camera shake is the friendly case. A hand moving during a slow exposure smears each point of light along a path, which is a convolution (2026) with a known-ish shape. Reverse the shape and much of the edge structure returns. Motion blur (2026) from a moving subject behaves the same way in the region that moved.

Missed focus is worse. An out-of-focus lens spreads every point into a disc, and overlapping discs destroy fine structure rather than displacing it. Heavy JPEG compression is worse again: the file no longer contains the high-frequency data, only a blocky approximation of it, and there is nothing under the blocks to recover.

So what do the tools do about the second and third cases? They generate. A model trained on millions of sharp photographs writes in the edges that a picture like this one usually has. That is genuinely useful for a portrait you want to print and genuinely wrong for anything used as evidence, and the difference is worth holding on to for the rest of this piece.

Open the file at 100 percent, pick a hard edge, and ask three questions in order. The whole diagnosis takes about fifteen seconds and it decides which of the later sections applies to you.

1. 1

   Is the smear pointing somewhere?

   Every highlight stretched along the same axis, the same distance, means the camera moved. Point lights become short lines. This is the recoverable case.

2. 2

   Is the subject soft while something else is sharp?

   If the railing behind the face is crisp and the face is not, focus landed in the wrong place. Nothing in the frame was smeared; the wrong plane was chosen.

3. 3

   Are the edges blocky and ringing?

   Eight-by-eight squares in flat areas, halos beside high-contrast lines, colour bleeding into skin. That is compression, usually from a file that has been re-uploaded and re-saved several times.

A fourth answer exists and it is the one nobody wants: the photo is fine and it is simply small. A 400-pixel avatar viewed at poster size is not blurred, it is enlarged. That case belongs in the low-resolution section below, and it responds to a completely different treatment.

Shake is the case where a single pass gets you most of the way, because the tool has real structure to work with rather than an empty field to fill.

Here is one run in the AI photo upscaler on a dusk street scene shot handheld. The composer is short on controls by design: the tool chip, a Photo box, a Prompt box, five ratio chips and three resolution chips. No sharpening slider, no radius, no threshold.

Both runs on this page went through GPT Image 2, which is the model sitting behind the image tools rather than a separate sharpening engine. That is worth knowing because it explains the behaviour: an unblur photo pass here is a generative rewrite constrained by your source, not a filter applied to it.

Two habits pay off on shake. Keep the aspect ratio the same as the original so nothing gets recomposed while you are trying to fix sharpness. And check the result against the source at 100 percent rather than side by side at fit-to-window, because a shrunk preview hides both the win and the damage.

This is where the answer to how to fix blurry photos stops being technical and starts being editorial. There is no true version of the eyelashes to recover, so the model supplies a plausible set. On a portrait the result is often lovely and often not quite the same person around the mouth and the eye corners.

• **Say what must not change.** The Prompt box is the only place to protect features. Naming the face, the glasses or the printed text on a sign is what keeps the pass from redesigning them.
• **Prefer one strong pass to three weak ones.** Each additional round treats the previous round's invention as evidence. Skin goes waxy on the second pass and plastic on the third.
• **Compare at the eyes, not the background.** A model will nail brickwork and quietly rebuild an iris. Brickwork is not what anyone is looking at.
• **Disclose when it matters.** For journalism, insurance, legal or listing photography, a reconstructed face is a fabricated detail. A tighter crop is often the defensible answer instead.

A sharpened portrait is a new photograph of an old moment. Useful on a wall, risky in a file that has to be true.

— Maya Chen

The phrase people actually type is make blurry picture clear, and nine times out of ten the picture is not blurred at all. It is small. Enlargement is a separate problem from sharpening, and the order matters. Repair or clean first, then enlarge, because image scaling (2026) multiplies whatever is already in the frame, artefacts included.

The workbench has no 2x or 4x button. Output size comes off the resolution chips instead, so a second run at 2K is how you ask for a bigger file. Here is that second run, this time on an indoor selfie rather than the street scene.

There is a ceiling and it arrives sooner than people expect. Past roughly four times the original linear size, extra output pixels stop describing the subject and start describing the model's idea of a subject. If the source is a 300-pixel-wide social avatar, no resolution chip in any product rescues it; the honest move is to find the original file.

Free and watermark-free are two separate questions, and several products answer yes to the first only. Everything in this table was read off the vendor's own pricing page on 7 September 2026.

Worth adding: nothing in the image generator reserves credits for a particular tool, so you can run the same file through sharpening once, then repair, then colour, and compare all three before deciding which one you actually needed.

Start with what is already installed. Both platforms ship something, and both stop at the same wall.

A word on searches for how to unblur images on a phone that end at a store listing: the single-tap apps are fine and they are also the least controllable option, because there is no place to tell them what to leave alone. On a face, that is the setting that matters most.

Run this list at 100 percent before you export anything. Every item shows up at full size and disappears in a thumbnail, which is exactly why over-processed images keep getting published.

• **Halos.** A pale outline hugging every dark edge against a bright background. Roof lines against sky give it away first.
• **Wax skin.** Pores gone, a smooth gradient where texture should be, and a faint sheen across the forehead.
• **Plastic hair.** Individual strands merged into ribbons that all curve the same way.
• **Invented text.** Signage, book spines and number plates that now read as almost-words. The clearest proof the output is generated rather than recovered.
• **Crunchy noise.** Grain that has been sharpened along with the subject, so the noise itself has edges.

If two or more of those appear, the answer is not a gentler pass on the same file. It is a better source: the original rather than the messaged copy, the RAW rather than the export, the camera roll rather than the chat thread. Most of the time it exists, and one message asking for it beats an hour of processing.

One habit worth breaking: 124 of 143 external image jobs in the August 2026 workbench figures came out square. Forcing a 3:2 frame into 1:1 hands the model a recomposition task stacked on top of the sharpness task, and recomposition is where faces drift.