Vision Harnessing Agent for Open Ad-hoc Segmentation
The episode discusses 'Vision Harnessing Agent for Open Ad-hoc Segmentation' (VASA), a training-free approach that uses an iterative visual construction process. Hosts discuss how VASA plans segmentation by selecting tools like ADD, REMOVE, or REPLACE based on candidate masks, using a persistent working mask and state management to reason through visual edits until the goal is met. The paper shows VASA outperforms SAM3 Agent.