基于视觉先验驱动的可控文本到图像生成技术综述OA

While text-to-image generation technologies,epitomized by diffusion models,have achieved groundbreaking advancements,relying solely on a single text modality struggles to precisely control the spatial structure and appearance style of images.Visual prior-driven controllable generation technologies can effectively solve this problem.Centering on the technological evolution path from"single control"to"multi-dimensional synergy,"this paper systematically reviews the visual prior-driven controllable image generation system in the latent space,covering both the classic U-Net architecture and the emerging Diffusion Transformer(DiT)paradigm.Building upon this,a core analytical framework of"Prior Injection-Feature Conflict-Synergistic Resolution"is constructed:from the three dimensions of geometric structure,layout positioning,and appearance style,the representation and injection mechanisms of single visual priors are analyzed;furthermore,synergistic resolution strategies based on instance perception and dynamic adjudication,attention mechanism regulation,as well as timestep scheduling and feature layer disentanglement,are critically discussed to resolve the interactive interference and feature conflicts between structure-layout,layout-appearance,and structure-appearance during multi-condition fusion.In addition,common datasets and quantitative evaluation systems are summarized,the challenges and emerging fine-grained metric paradigms facing multi-dimensional controllability assessment are deeply analyzed,the scenario-based applications of image generation technologies are reviewed,and the future development trends of this field are prospected.