Building an Offline Evaluation Set for Outpainting

Your team has a solid offline evaluation pipeline for inpainting (removal and generate-with-prompt): synthetically mask a region of a real, held-out image and compare the model's fill against the known original pixels. A PM asks you to build the equivalent evaluation for outpainting before the feature ships.

1. Explain why the inpainting evaluation approach can't be applied to outpainting unmodified, and what's missing.
2. Propose a concrete construction for an outpainting offline evaluation set that provides genuine ground truth.
3. Even with your proposed ground-truth set, what aspect of outpainting quality would still require human evaluation rather than an automated metric, and why?