EVENT DOSSIER
On September 7, 2026, arXiv cs.CV published the paper “Compositional Reward Models for Conditional Medical Image Generation”. This study proposed a framework based on compositional reward models, aimed at improving the quality and controllability of conditional medical image generation. The method uses a specific reward function to guide the generation process, enabling it to synthesize medical images that meet clinical requirements more accurately based on the input conditions.
提出 PRISM 框架以解决高质量标注医疗数据获取成本高、现有生成模型难以捕捉细粒度属性及语义一致性问题。该框架采用组合奖励模型（CRM），将图像质量分解为从细到粗的多个验证阶段，涵盖低层属性、结构对齐及高层语义保真度，并通过分层约束传播机制确保低级缺陷解决后再积累高级奖励。研究在 PanNuke、CeDeM 和 ISIC 三个数据集上评估了该框架，结果显示使用 PRISM 生成数据训练下游模型，相比最接近基线分别实现了 mDice 提升 2.3%、平均相对误差降低 8.5% 以及 F1 值提升 5.9%。