Recover a softly blurred photo by inverting the blur. A photo that missed focus, or moved during the exposure, is the sharp scene convolved with a small kernel: a disc for a missed focus, a short line for camera shake. Deconvolution solves for the scene given that kernel. On a blur of a few pixels the result is a genuinely sharper picture, with text readable again and edges back to a single line.
Choose the kernel
Blur type names the physical cause. Lens is a defocus: every point became a disc, and the slider sets its radius in source pixels. Motion is a smear along one direction: the slider sets its length and Angle sets the direction, 0 for horizontal shake and 90 for vertical. Soft is a Gaussian haze with no single cause, which fits a picture that was upscaled, denoised too hard, or scanned slightly out of focus.
The size is the setting that matters most, and the picture tells you when it is right. Too small and the result is only a little sharper. Too large and edges grow a halo, then a second ripple. The correct value sits just before the halo appears. For Motion, press Estimate motion first: it reads the blur length and direction from the crop’s spectrum and fills both sliders when it finds a clear peak.
Reading the split preview
The preview is a crop at one source pixel per screen pixel. Everything to the left of the line is the original, everything to the right is the result, and the line drags. Judge fine text, an eyelash, a brick edge, not a smooth sky, where no method can show anything.
The navigator in the top corner shows the whole picture with the crop marked. Drag it onto the part that matters, usually the subject’s eyes or the text you need to read, because that is the region the settings should be tuned on.
Strength
The solver minimises the difference between the blurred estimate and your photo, balanced against a total-variation prior that keeps the result piecewise smooth. Strength moves that balance. At the low end the prior dominates: noise stays down and some softness remains. At the high end the data dominates: more detail comes out, along with sensor noise, JPEG blocks and ringing around high-contrast edges. Most photos sit well around the middle; a clean low-ISO file from a camera can go higher, and a phone JPEG shot at night should go lower.
The processing runs on linear-light luminance only, and colour is scaled to match. That keeps chroma from ringing and avoids the coloured fringes a per-channel deconvolution produces.
Limits
Deconvolution inverts a blur; it does not create information. A blur larger than roughly 15 pixels has erased whole bands of the spectrum, and raising Strength there amplifies noise rather than recovering anything. A blur that varies across the frame, such as a portrait with a soft background or a rotation shake, has no single kernel to invert. A heavily compressed JPEG shows its 8 by 8 block grid once sharpened, because the grid is real detail as far as the maths is concerned.
A blurred face stays a blurred face. The search results that promise a sharp face from a hopeless one use a generative model that invents a plausible one, which is a different job with a different honesty. This page recovers what the file still holds, which on a mildly soft photo is more than most people expect.
Export
Download processes the full-resolution image in overlapping tiles with the same kernel and Strength as the preview, blends the tile edges, and writes the source format. Output is a re-encoded file, so EXIF data from the original is not carried over.