Fitting an 8x Restoration and a Phone NPU Model Under Memory Limits

Two separate engineering complaints land in the same week: (1) the cloud restoration service is running out of GPU memory on a batch of unusually large, high-resolution scanned photos submitted for 8x restoration, and visible seams appear in the output where the team's existing quick-fix (processing the image in fixed non-overlapping blocks) meets block boundaries; (2) the on-device model, which hits its latency budget fine on newer phones, is timing out on older, lower-end phones that lack the NPU delegate the model was primarily tuned for.

1. Diagnose the seam problem precisely — why does non-overlapping block processing produce visible seams, and what is the fix?
2. Propose a concrete plan for the older-phone timeout problem, using at least two named edge-deployment techniques from this case study.
3. Explain why "just always run the cloud model" is not a valid fallback for problem (2), tying your answer back to this case's central design fork.