The three kinds of blur — and what's recoverable
Soft focus is the best case: the detail is spread out but still present, so the model has strong evidence for where every edge belongs. These shots often come back looking like the focus simply landed. Camera shake is the middle case — the directional smear is harder to invert, but small amounts recover well.
Heavy motion blur is the honest limit. When a subject moved enough to paint itself across many pixels, the original detail is gone and the model has to invent a plausible replacement. The result can look good, but treat it as a reconstruction, not a recovery — especially on faces you know well.
How AI deblurring actually works
The Enhance model learned from enormous numbers of paired sharp and degraded images, so it carries a statistical memory of what real-world detail looks like under blur: how a jawline smears, how out-of-focus text bleeds, how compression eats fabric texture. Given your photo, it predicts the sharp version most likely to have produced it.
This is why the failure mode is graceful rather than ugly. Where the evidence is strong the output is precise; where the evidence is thin the model falls back to plausible texture. You get a usable image either way — you just shouldn't rely on invented regions for factual detail.
Getting the most out of a rescue
Always start from the best original you have. The version in your camera roll beats the one that went through a messaging app, and a full-size export beats a screenshot of it. Every generation of compression before the deblur pass removes evidence the model could have used.
After enhancing, zoom to 100% on the areas that matter — eyes, text, product edges — before you call it done. If the shot is headed for print, run Upscale after Enhance, in that order: deblur first so the enlargement magnifies clean detail instead of soft detail.