Multiscale feature generation and innovative design method for paper-cut patterns using diffusion models with semantic constraints
The Changbaishan Manchu paper-cutting is not only an important cultural symbol but also a key visual resource for the design of creative cultural products. In response to the problem of the lack of engineering modeling and controllable generation mechanisms in the existing development, this paper constructs a standardized paper-cutting dataset including collection, preprocessing and annotation, introduces multiscale graphic feature vectors and complexity indices to quantitatively represent shape, topology and decorative style; further designs a Manchu cultural semantic ontology and knowledge base of composition rules, converting the compliance of image combination into computable semantic constraints. On this basis, a conditional diffusion generation framework of "multiscale condition encoding + semantic constraint guidance" is proposed: with the denoising diffusion model as the core generator, through multiscale feature pyramid injection and semantic token cross attention to achieve joint control of structure and image, and in the sampling stage, introducing guidance terms of rule compliance and style deviation to suppress invalid samples that violate traditional composition. For application scenarios such as packaging, a layout optimization objective function and a cross-carrier "innovation degree - style consistency" joint control model are constructed, forming an interactive parameter-driven human-machine collaborative design workflow, and the generation and design support system is implemented under the browser/server architecture. Experiments show that the style consistency of the system has increased from 0.82/0.80 to 0.90, the novelty has increased from 0.42/0.55 to 0.78, user satisfaction has increased from 3.7/3.9 points to 4.6 points, the average design time has decreased from 260s, 180s to 110s, verifying the comprehensive advantages of the proposed method in innovation expression quality and design efficiency, and providing a reusable engineering path for the multiscale controllable generation of paper-cutting patterns and intelligent cultural product development.
Paper
Full text
Multiscale feature generation and innovative design method for paper-cut patterns using diffusion models with semantic constraints
Semantic Scholar · Art · 2026
Abstract
The Changbaishan Manchu paper-cutting is not only an important cultural symbol but also a key visual resource for the design of creative cultural products. In response to the problem of the lack of engineering modeling and controllable generation mechanisms in the existing development, this paper constructs a standardized paper-cutting dataset including collection, preprocessing and annotation, introduces multiscale graphic feature vectors and complexity indices to quantitatively represent shape, topology and decorative style; further designs a Manchu cultural semantic ontology and knowledge base of composition rules, converting the compliance of image combination into computable semantic constraints. On this basis, a conditional diffusion generation framework of "multiscale condition encoding + semantic constraint guidance" is proposed: with the denoising diffusion model as the core generator, through multiscale feature pyramid injection and semantic token cross attention to achieve joint control of structure and image, and in the sampling stage, introducing guidance terms of rule compliance and style deviation to suppress invalid samples that violate traditional composition. For application scenarios such as packaging, a layout optimization objective function and a cross-carrier "innovation degree - style consistency" joint control model are constructed, forming an interactive parameter-driven human-machine collaborative design workflow, and the generation and design support system is implemented under the browser/server architecture. Experiments show that the style consistency of the system has increased from 0.82/0.80 to 0.90, the novelty has increased from 0.42/0.55 to 0.78, user satisfaction has increased from 3.7/3.9 points to 4.6 points, the average design time has decreased from 260s, 180s to 110s, verifying the comprehensive advantages of the proposed method in innovation expression quality and design efficiency, and providing a reusable engineering path for the multiscale controllable generation of paper-cutting patterns and intelligent cultural product development.