DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing
DARS tries to tell whether a failed image edit came from the planning step or the rendering step.
The paper targets planner-renderer image editing systems, where final-image rewards alone do not show which module needs improvement. Its method uses multiple plans and renders to route credit between modules and adjust training difficulty. Inside the planner, it structures reasoning into four fields so rewards can be pushed back to more specific parts of the output. The authors report stronger results than a matched Joint RL baseline on five benchmarks, especially for reasoning-heavy edits. ArXiv · AI/CL/LG's note
The paper targets planner-renderer image editing systems, where final-image rewards alone do not show which module needs improvement. Its method uses multiple plans and renders to route credit between modules and adjust training difficulty. Inside the planner, it structures reasoning into four fields so rewards can be pushed back to more specific parts of the output. The authors report stronger results than a matched Joint RL baseline on five benchmarks, especially for reasoning-heavy edits. ArXiv · AI/CL/LG's note
score 5