Dynamic Visual Design System Driven by AIGC: From Parameterized Generation to User Preference Learning

The current visual design system has problems such as insufficient personalized adaptation ability, dependence on static parameters in the generation process, and difficulty in dynamically responding to users’ deep preferences. This paper proposes a dynamic visual design system driven by AIGC (Artificial Intelligence Generated Content). This paper first constructs a multi-scale parameterized generation module based on conditional variational autoencoder, which achieves structured control of visual attributes such as color, shape, and composition through a layered and decoupled latent variable space; and further introduces a deep reinforcement learning module based on the Double Q network, which models the user interaction behavior sequence as a state action decision process, quantifies user preferences through reward signals, and dynamically adjusts the generated parameters. The two modules form a closed-loop optimization through end-to-end gradient collaboration, enabling the system to continuously adapt to user needs based on real-time feedback. The experiment showed that the system achieved an average user preference matching degree of 0.885, an average generation efficiency of 93.7 seconds, and a sample diversity score of 0.905 in 10000 user interaction tests. Its precise adjustment ability to increase user ratings from 5.2 to 9.4 within four rounds of interaction was verified through case studies. This system achieves a paradigm shift from standardized output to personalized intelligent design.