Diagnosing a Model That Fails on Real User Masks

Your inpainting model's offline benchmark (built from randomly generated rectangular and free-form splotch masks) shows strong results. In production, two specific complaint patterns emerge: (a) "remove this person" requests, where users draw a mask that roughly traces the person's silhouette, frequently produce poor fills with visible ghosting; (b) outpainting requests to extend a photo's aspect ratio for an ad banner frequently produce a visibly different lighting style in the extended region compared to the original photo.

1. What is the most likely root cause common to both complaint patterns, tracing back to the training data pipeline?
2. Propose a concrete fix for the training-data mask distribution that addresses both patterns, and explain the mechanism by which it would help each one.
3. A teammate proposes simply generating more random splotch masks — 10x more training examples with the existing mask-generation approach — as the fix. Evaluate this proposal.