Full resolution in, uncompressed out.
When your hand moves while the shutter is open, every point of light is recorded as a short streak instead of a point. Sharpening cannot put that back, it only raises contrast on whatever survived. This uses a neural network trained on 7,400 pairs of the same scene, captured once shaken and once still, so it has learned what the smeared detail is supposed to look like and reconstructs it.
The pairs all come from one phone, so photos from other cameras, shots that were out of focus rather than shaken, and very dark scenes may not improve.
Photos are processed in memory and never stored.