Composite-Semantic Control for NEV Styling Using Bayesian Evidence Updating and LoRA-Tuned Diffusion Models

With the rapid iteration of New Energy Vehicle (NEV) styling, design intent has evolved from single adjectives to coupled semantics such as “technological-elegant” and “dynamic-steady.” Existing text-to-image diffusion pipelines typically rely on prompt concatenation or heuristic weight tuning, which often leads to token competition, cross-sample drift, and limited controllability under composite constraints. This paper proposes a composite-imagery driven framework for NEV styling generation via Bayesian posterior control. First, composite imagery is modeled as a probabilistic semantic weight vector, and CLIP-based similarity is used as visual evidence to define a likelihood surrogate for Bayesian updating, producing a semantic posterior. Second, the posterior is treated as a continuous control signal; high-dimensional semantics are clustered and compressed into interpretable axes to enable slider-based interactive control. Third, an NEV-specific LoRA is trained on Stable Diffusion, together with soft-label supervision to inject domain priors and improve structural coherence. Comparative evaluations with human experts and a Multimodal Large Language Model (MLLM) indicate that the proposed method outperforms prompt-weighting baselines, yielding consistent improvements in semantic alignment and structural completeness. The proposed method provides an evidence-driven, controllable generation mechanism for early-stage automotive styling tasks.