Agent Banana：基于代理思维和工具的高保真图像编辑

This study addresses instruction-based image editing within professional workflows and identifies three persistent challenges: (i) editors often over-edit, modifying content beyond the user's intent; (ii) existing models are predominantly single-turn, while multi-turn edits can compromise object fidelity; and (iii) evaluation at around 1K resolution is misaligned with real workflows that typically operate on ultra high-definition images (e.g., 4K). Agent Banana is proposed as a hierarchical agentic planner-executor framework for high-fidelity, object-aware, deliberative editing. It introduces two key mechanisms: (1) Context Folding, which compresses long interaction histories into structured memory for stable long-horizon control; and (2) Image Layer Decomposition, enhancing the flexibility and precision of editing.