DynRefer: Delving into Region-level Multimodal Tasks via Dynamic Resolution | Read Paper on Bytez