Generalization to Larger Numbers of Reference Images

Determine whether the public MSR visual workflow generalizes reliably to more than five reference images.

Background

The public MSR workflow exposes an interface for up to five reference images, whereas the documented training data include only a very small number of examples with six to eight images and do not evaluate generation quality at those counts. The paper therefore leaves unresolved whether the slot-aware conditioning scheme remains effective when used with more than five references.