Why the Bicubic-Trained Model Fails on Real Phones

Your team trains a super-resolution model on pairs constructed by taking clean high-resolution images and generating low-resolution inputs via a single bicubic downsample. The model achieves excellent metrics on a held-out test set built the same way. In production, applied to real phone photos, users report the model barely changes anything on noisy or heavily-compressed images, and occasionally adds strange artifacts on photos that have been through several rounds of messaging-app re-compression.

1. Explain why strong performance on the bicubic-pair test set doesn't predict production performance here.
2. Describe, concretely, the fix (high-order degradation modelling) and why it's expected to generalize better, referencing what makes real photo degradation different from a single bicubic downsample.
3. Your team has budget to collect a modest amount of genuinely real (non-synthetic) paired data via burst captures at two zoom levels. How would you use it, given it can't scale to full training volume?