Objective: Architectural free-form surfaces often rely on nonuniform rational B-splines for delineation, posing significant challenges for grid partitioning owing to their diverse and irregular configurations. Prevailing grid partitioning methods are tailored to specific free-form geometries, leading to a lack of universality in extant explicit programming algorithms because of their specificity. Structural design, including grid partitioning, largely depends on the empirical knowledge and intuitive judgment of designers. As interdisciplinary collaboration and efficient design processes escalate, the need for accuracy and speed has increased. To alleviate the intrinsic limitations of explicit programming and mitigate overdependence on the designer acumen, this paper proposes the use of a generative adversarial network model to elucidate and integrate the logical correlation between free-form surfaces and their corresponding grid structures. This approach enables the generation of grid structures from free-form surfaces as inputs to the generative adversarial network model. Methods: The process starts with a preprocessing regimen for free-form surfaces. To fit the two-dimensional input and output framework of the generative adversarial network model, a self-developed algorithm generates curvature and height cloud maps representing the free-form surface, which are used as inputs to the generative adversarial network model. The pix2pixHD model is modified to allow both curvature and height cloud maps to be input simultaneously into the generator and discriminator. These cloud maps are then fed into the grid generative adversarial network(GridGAN) model, which has been pretrained and validated to derive grid partitioning outputs. In the postprocessing phase, the two-dimensional grid data is transformed into three-dimensional grid structures by extracting nodal points and their topological relationships from the grid layout. This information is subsequently projected into three-dimensional space. The effectiveness of the proposed method is demonstrated through a comparative analysis with two existing explicit programming grid partitioning algorithms (one quadrilateral and one triangular). Multiple examples are used to evaluate the generative design approach, employing evaluation metrics based on grid geometric properties such as rod length factor and shape quality factor. Results: The case studies indicated that the intelligent free-form grid partitioning method proposed in this paper performed comparably to the triangular and quadrilateral grid partitioning algorithms used in explicit programming. The maximum relative error recorded was 2.908% for the mean rod length and 1.133% for the shape quality factor, both of which fell within acceptable limits. Conclusions: These findings confirm that the proposed approach achieves grid partitioning outcomes comparable to those of various explicit programming algorithms. It effectively handles free-form surfaces with diverse shapes. This research introduces a methodological perspective in architectural design and establishes a robust foundation for future research and applications. It has the potential to catalyze the evolution of design practices toward greater efficiency and intelligence.