A LLM-Based Dynamic Deduction Framework for Early-Stage Visual Design: Integrating Cultural Conventions and Multimodal Verification

With the rapid advancement of artificial intelligence (AI) technology, its application in the field of design has significantly enhanced design efficiency and innovative potential. However, its capabilities in dynamic deduction and rationality verification during the early design stages remain distinctly limited. Existing studies predominantly focus on the automated generation of design outcomes, yet fail to provide systematic support for explicating design-path logic, cross-dimensional scenario rehearsals, and human–machine collaboration. In light of these limitations, this article introduces a dynamic deduction method tailored for early-stage visual design based on large language models (LLMs). The approach involves the construction of an adversarial prompt framework and the integration of multimodal analysis techniques, thereby transforming implicit design experiences (such as visual weight balance) into quantifiable and verifiable indicators. Additionally, a “spatial perception simulation” prompt chain is developed to realize the deduction of visual impact attenuation in real scenarios. The study constructs a dynamic design dictionary containing over 200 professional terms, thereby supporting nonprofessional users in precisely articulating their design requirements. Meanwhile, the cross-media consistency of design schemes is verified rigorously through multimodal data fusion. The experimental results provide preliminary evidence that the proposed method can improve early-stage design efficiency and support more explicit reasoning, scenario-aware evaluation, and cross-role communication under the tested conditions.