计算机应用研究2026，Vol.43Issue(9)：2589-2596,8.DOI:10.19734/j.issn.1001-3695.2026.02.0023
交错模态思维链奖励机制的细粒度图像生成
Fine-grained image generation via interleaved multimodal chain-of-thought reward
摘要
Abstract
Fine-grained semantic alignment remains a major challenge for current text-to-image generation models,especially in compositional scenarios involving attribute binding,numeracy,and spatial structure.Existing methods often rely on specific model architectures and high-cost training or inference,which limits stable semantic correction and unifies optimization across paradigms.To address this issue,this study proposed ICOTR,a unified semantic alignment reward fine-tuning framework based on an interleaved multimodal chain-of-thought reasoning mechanism.ICOTR used the image-text understanding capabili-ty of multimodal large language models to build adaptive multi-dimensional semantic reward feedback,and applied consistent semantic guidance to both diffusion and autoregressive models.Experiments on the high-semantic-complexity benchmark T2I-CompBench++show that ICOTR improved the overall performance of SDXL from 0.411 4 to 0.496 5 and that of Show-o from 0.381 5 to 0.520 2.The results indicate that ICOTR consistently outperforms existing methods in overall performance and se-veral key subtasks,and effectively improves fine-grained image generation across different paradigms.关键词
细粒度文生图/思维链/奖励微调/多模态大模型Key words
fine-grained text-to-image generation/chain-of-thought/reward fine-tuning/multimodal large language model分类
信息技术与安全科学引用本文复制引用
郑凯,周远远,徐林莉..交错模态思维链奖励机制的细粒度图像生成[J].计算机应用研究,2026,43(9):2589-2596,8.基金项目
国家自然科学基金资助项目(62276245) （62276245）
