Deblur an Image — Pro, on our server
Blurring is a convolution, and convolution loses information on purpose. Each output pixel is a weighted average of a neighbourhood of input pixels, the weights described by a point spread function — a streak for camera shake, a disc for defocus, a soft bell for atmospheric haze. Averaging is a one-way door: once you know only the average, you cannot say which of the many possible neighbourhoods produced it. Fine detail is where this bites first, because the highest-frequency structure is precisely what an average flattens towards nothing, and no arithmetic recovers a value that has been multiplied by zero.
Classical deconvolution needs two things a real photograph does not give you. It needs the point spread function, which you do not know, and it needs a clean signal, which you do not have — every file carries noise, and inverting a blur amplifies exactly the frequencies where noise dominates. Estimating both the blur and the scene at once is blind deconvolution: badly underdetermined, and prone to ringing ripples around high-contrast edges that look worse than the blur did. That failure is the reason the Cloud HD enhancer uses a trained model instead of an inversion.
A learned prior substitutes knowledge for the missing measurements. The model has seen enormous numbers of blurred-and-sharp pairs, so it does not have to solve for the blur; it proposes the sharp photograph that both explains your pixels and looks like the kind of thing photographs look like. That is why results are so much more stable than deconvolution and why the honest word for the output is reconstruction. It also sets the ceiling: where the blur is severe, the prior is doing most of the work, and the fine detail you see is inference rather than record. The practical consequences for the two common cases are on unblurring a shaky image and sharpening a soft photo.
Give the model one clean problem, and know where it solves it. If you are converting from another format first, export losslessly with the on-device image converter so compression damage is not a second fault tangled with the first — that conversion, like nearly everything on MiniPx, happens inside your browser. Cloud HD is the exception and is Pro-only and opt-in: the image 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. The model that ran is named on the result — RestoreFormer++ for now; the model can change, and you can always see which one ran.