Enhancing Creative Expression Skills in Digital Media Education Using CLIP-Driven Multimodal Generation Systems

Current digital media education lacks a quantitative assessment method for creativity that simultaneously considers semantic fidelity, aesthetic quality, and stylistic diversity. This paper proposes a CLIP-driven multimodal generation and evaluation framework for improving creative expression skills. In the generation phase, the CLIP text encoder extracts semantic vectors from student prompts and feeds them into a diffusion model as conditional inputs. During sampling, embeddings of intermediate images and corresponding prompts are compared to guide semantic correction. In the assessment phase, CLIPScore is used to quantify semantic fidelity, a neural image assessment model outputs aesthetic quality scores, and principal component analysis is performed on CLIP visual embeddings from multi-round generated images to compute style dispersion through the covariance matrix trace. These three indicators are weighted to form a comprehensive creative-expression score. Propensity score matching and difference-in-differences estimation are further applied to identify the net causal effect of the system intervention. Experiments involving digital media students show a median semantic fidelity of 0.631, mean visual expressiveness of 0.774, style exploration breadth of 0.629, and average comprehensive creative-expression score of 0.782. The framework enables three-dimensional quantitative assessment and causal evaluation of creative expression in digital media education.