Two long reads
The front page has room to state a position. These have room to argue one. Both exist because the honest answer to the question underneath a search here is longer than a control label.
Pixels stay in this tab; only the file you save leaves it.
Start with whichever question you arrived with
If the question is whether any of this works, the first one is the place to start: it is about what a trained network has actually learned, which failures are inside that and which are outside it, and the specific way a reconstruction goes wrong when it goes wrong.
If the question is what it costs, the second one answers it in bytes, in seconds and in storage, including the part most pages leave out — what is written to your device and how to get rid of it.
- What a deblur model can reconstructLearned priors against measured smears, why the GoPro corpus makes shake learnable and defocus not, and how to read a result sceptically.What can be reconstructed
- What ninety-two megabytes costsWhat is fetched, from where, what is written to your device, how to clear it, and what a slow phone should expect from the arithmetic.What 92 megabytes costs