Evaluation of mask–extent decoupling and depth-controlled resizing

Establish an evaluation protocol that directly measures mask–extent decoupling and depth-controlled resizing in local image-insertion backends for relative-scale correction.

Background

Rescale performs localized object resizing and reinsertion using an editable mask, a completed background, an object reference crop, and optional depth conditioning. The specialized depth-aware backend is intended to prevent the resized object from simply filling the entire editable mask, thereby decoupling the spatial extent reserved for editing from the final extent of the inserted object.

The paper reports that existing crop-level metrics primarily measure visual consistency and no-reference quality, not whether an insertion backend can use a large editable mask while preserving the intended object size and applying depth-controlled resizing. The authors therefore identify the development of a targeted evaluation protocol as an unresolved methodological problem for assessing scale-aware insertion.