Fix an Out-of-Focus Photo — Pro, on our server
Defocus spreads each point of the scene into a disc, and the disc is the problem. A lens focuses one distance onto the sensor exactly; everything nearer or farther arrives as a small circle rather than a point, and the further off the distance, the wider the circle. Photographers call it the circle of confusion, and the band of distances where it stays small enough to look sharp is depth of field. That geometry is why a focus miss looks different from camera shake: the softness spreads outward in every direction rather than along a track, and it varies with distance across the frame instead of being uniform. The Cloud HD enhancer works with what the discs left behind.
A disc is harder to undo than a line. Shake mixes each pixel with its neighbours along one direction, so some independent information survives across the perpendicular one. Defocus mixes each pixel with everything around it, which destroys information faster for the same visual amount of softness. Practically, that means the recoverable window here is narrower than on the shaky-image page, and the mathematics of why sits on the page about deblurring an image, which explains the inverse problem that both cases really are.
Faces are the part most worth trying anyway. Even where general recovery is limited, the face pass runs a model trained on faces alone, so a moderately soft face often gains more than the wall behind it. The restored face is checked against the one that arrived and pasted back under a face-shaped mask, so hair and clothing keep your pixels. Before you upload, crop to the subject with the on-device crop tool — and if you are still unsure whether your fault is focus, motion or simply too few pixels, the blur triage page sorts that out first.
What it costs and where it runs. One pass is a single general Cloud HD run — 150 a month on Pro Monthly and 200 on Pro Annual, resetting on the 1st in UTC with no rollover and a ceiling of 30 runs an hour; prices exclude tax. Cloud HD is the one part of MiniPx that leaves your device, and it says so plainly: the photo is uploaded to our server for that run and deleted when the run finishes, and in all cases within 24 hours, processed in the United States, encrypted in transit, and never used to train anything. RestoreFormer++ is the model running today, and every result names it.