Figure 2: Overview of GRASP. Real-ISR is cast as visual in-context completion. Local visual
evidence comes from the diptych, global image context from the Vision-Context Bridge over CLIP,
and semantic prompts from LLaVA captions during training and agent refinement during inference.