Researchers have developed a new reinforcement learning framework called DARS for instruction-based image editing. The framework assigns credit to both the planner and renderer stages of the process, allowing for more efficient training with rewards. This is achieved through multi-plan rollouts that estimate reward variability between plans and within plans, as well as a structured reasoning output that enables localized supervision. Experiments show that DARS outperforms a baseline method on five benchmarks, particularly for edits that require complex reasoning. --- Why it matters: This matters to researchers in AI because it addresses the issue of inefficient training with only final-image rewards, allowing for more effective optimization of instruction-based image editing systems. Source: https://arxiv.org/abs/2608.20161