A novel generative design approach based on conditional diffusion model and multi-criteria decision-making under biomorphological and emotional constraints

Biomorphological product development increasingly requires the simultaneous consideration of biological inspiration and users’ emotional preferences. However, existing generative design methodologies often struggle to coordinate heterogeneous visual and semantic constraints, while the evaluation and selection of generated concepts remain challenging under uncertain decision environments. To address these challenges, this study proposes a novel generative design approach based on a conditional diffusion model and multi-criteria decision-making (MCDM) under biomorphological and emotional constraints. The proposed framework integrates controllable generation and reliability-aware evaluation within a unified design workflow. To facilitate the development and validation of the presented approach, a biomorphological furniture design dataset containing 1197 annotated samples with five-dimensional emotional labels and multi-color variants is constructed. Building upon the developed dataset, a gated multi-layer perceptron attention adapter (GMAtt-Adapter) is designed to mitigate conditioning conflicts through adaptive gradient balancing, nonlinear feature mapping, and variational latent representation learning. Furthermore, an MCDM method, termed Z-CqFrFS-VIKOR, is formulated by integrating Z-number and cubic q-rung fractional fuzzy sets into a possibility-degree-based VIKOR framework, enabling the simultaneous incorporation of evaluation uncertainty and expert confidence during design selection. Experimental results demonstrate that GMAtt-Adapter reduces the Kernel Inception Distance by more than 40% and shortens training time by over 50% compared with representative conditional generation models. In addition, the proposed Z-CqFrFS-VIKOR method demonstrates consistent ranking performance, the lowest fuzzy entropy among competing methods, and robustness under parameter variations.