DARS Framework for Instruction-Based Image Editing via Dual-Level RL

DARS uses dual-level credit assignment to optimize planner-renderer pipelines in image editing. It employs multi-plan multi-render rollouts and a four-field structured reasoning output to better assign rewards to either the VLM planner or the diffusion renderer.