Work Out Which Blur You Have, Then Fix It

How it works

The image is scaled to about one megapixel and repainted at a moderate denoise with a fresh seed each run. The instruction asks for the edges the blur spread out to be reconstructed — faces, text, and fine structure named specifically — while holding the composition, every person's identity, and the depth of field of the original. The negative prompt argues against double edges, ringing outlines, oversharpened haloes, and invented text.

Depth of field is in the instruction for a reason. A photograph with a deliberately soft background should keep it. A tool that sharpens everything equally does not fix a blurred picture, it destroys a composed one.

Four blurs, ranked by hope

Camera shake is the most recoverable. The whole frame moves together along a short path, so every point is drawn into the same small streak. That path is a strong constraint: the damage is consistent everywhere, and the ordering of detail survives within each streak. Look at a small bright light in the frame — if it has become a short line or a squiggle, this is what you have, and it is the good news.

Subject motion is partly recoverable in appearance and not at all in fact. The background is sharp because the camera was still. The person is smeared because they were not. Their outline genuinely occupied several positions during the exposure, so there is no single true shape to return to.

Defocus is the hard one. When focus misses, each point of the scene is projected as a small disc rather than a point, and adjacent discs overlap and average. Averaging is not reversible: many different arrangements of detail produce the same blurred result, and nothing in the file says which one you had. A bright point becomes a round blob rather than a streak, which is how you spot it.

Lens softness is the least dramatic and the most treatable, because a slightly soft lens has recorded everything at low contrast rather than mixing it together. Conventional sharpening handles it and this page is overkill.

The most convincing wrong answer in imaging

This is the warning the page exists to give.

When a denoiser fails you get a mushy picture and you know it failed. When a deblurring pass fails you get a sharp, clean, plausible picture, and the failure looks exactly like success. There is no artifact to notice, no softness to distrust, nothing marking which features were measured and which were composed.

That matters most where blur is most common: a face, a number plate, a document in the background of a photograph. Each of those is a case where the reader wants a specific answer, and where a confident wrong answer is worse than an admitted absence.

So use the output for the family album freely, and never for identifying a person or reading a number that was not legible before.

Compare, do not admire

The practical habit is straightforward and most people skip it.

Open the original and the result at full size side by side, and look specifically at the parts you care about. Not whether the new one is sharper — it will be — but whether the same things are in the same places. A collar that changed shape, an eye that moved, a word that became a different word. Those are the signal that you are looking at a reconstruction, and they are visible in seconds once you look for them rather than at the overall improvement.

Publication gate

This page ships once the workflow has been run against a shake-blurred frame with point highlights, a defocused portrait, a photograph with a moving subject against a static background, and an image with deliberate background separation, confirming in the last case that the intended softness survives.