DARS Improves Instruction-Based Image Editing with Dual-Level RL.

Instruction-based image editing typically relies on a two-stage pipeline: a vision-language model (VLM) generates an edit plan, which a diffusion model then executes. A major challenge in training such systems is the inefficiency of using only final-image rewards. When an edit is poor, it's difficult to determine whether the planner or the renderer needs more optimization, and even within the planner's free-form reasoning trace, localizing the exact cause of failure is hard. To address this, the DARS (Dual-Level Credit Assignment RL with Structured Reasoning) framework has been developed. DARS introduces a reinforcement learning approach that performs credit assignment at two levels. Across the modules, multi-plan multi-render rollouts estimate reward variability to enable soft module routing and provide hardness estimates for an adaptive curriculum. Within the planner, DARS utilizes a four-field structured reasoning output, allowing for prefix-gated rewards and token-level advantage reweighting. This transforms outcome-level feedback into precise, localized supervision. Experiments across five benchmarks demonstrate that DARS significantly outperforms a Joint RL baseline, especially on complex, reasoning-intensive edits, by effectively localizing and applying feedback.