面向多主体语义解耦的图像生成框架研究OA

To address issues such as semantic confusion,attribute mismatch,and insufficient consistency in multi-subject and multi-attribute text-guided generation image with diffusion models,this paper proposes a semantic-guided multi-subject image generation method.First,a text-image preprocessing mechanism is designed,which enhances the model's ability to disentangle multiple entities and their attributes through part-of-speech segmentation and fusion with latent representations of reference images.Second,a hierarchical cross-attention mechanism is introduced,combining regional masks with entity-specific text to achieve precise semantic and spatial alignment.Finally,a dual-branch parallel denoising architecture is constructed,integrating structural priors from layout maps with textual semantics to improve structural consistency and semantic stability in the generated images.Experiments are conducted on the color,texture,and shape subsets of the T2I-CompBench benchmark using Stable Diffusion XL.Semantic consistency is evaluated using BLIP-VQA and VQAScore metrics,achieving scores of 69.79%,58.43%,52.97%,and 80.74%,respectively—representing improvements of 1.40%,1.09%,2.05%,and 2.04%over existing methods—demonstrating the effectiveness of the proposed approach in generating consistent and controllable images under complex text-image conditions.