About visual prompt object queries

[feivelliu]

For the interactive mode, following SAM, which employs prompt encoder to encode various visual prompts as location queries.
This has the limitation that it relies mainly on spatial information to guide the segmentation process and cannot query and segment the same visual target in another image.
Is it possible to similarly encode visual prompt as semantic queries like text to handle the segmentation of the same visual target across images?