CarvGPT: a multi-agent collaborative framework for cultural translation and digital innovation in traditional craft design

Creative design plays an important role in the contemporary development of traditional crafts. However, during the early conceptual design stage, translating cultural connotations into design solutions often relies on tacit experiential judgments and lacks a clear pathway, posing challenges for designers with limited domain experience. To address these challenges, we propose CarvGPT, a multi-agent collaborative framework for cultural translation and digital innovation in traditional craft design. During ideation, multi-agent collaboration is guided by a Semantic–Function–Craftsmanship–Form (SFCF) strategy to support the systematic translation of cultural connotations into design concepts. The SFCF-driven workflow integrates Requirement Analysis, Cultural Parsing, Translation-oriented Ideation, and Refinement and Decision-making. Through specialised role assignments, agents help designers understand cultural connotations and incorporate craft-specific considerations into concept development. Furthermore, LoRA fine-tuning is applied to adapt the diffusion model to the visual characteristics of Chinese boxwood carving. The results indicate that CarvGPT significantly improves the transparency of the ideation process and enhances users' perception of cultural connotations, while strengthening the representation of craft-specific visual characteristics in generated outputs. Grounded in a cultural translation framework, CarvGPT provides a structured and interpretable approach to digital innovation in traditional craft design.