Topology optimization is a powerful tool for producing an optimal free-form design from input mechanical constraints. However, in its traditional-density-based approach, it does not feature a proper definition for the external boundary. Therefore, the integration of shape-related constraints remains hard. It requires the experts’ intervention to interpret the generated designs into parametric shapes; thus, making the design process time-consuming. With the growing role of additive manufacturing in the industry, developing a design approach considering mechanical and geometrical constraints simultaneously becomes an interesting way to integrate manufacturing and aesthetics constraints into mechanical design. In this paper, we propose to generate mechanically and geometrically valid designs using a deep-learning solution trained via a dual-discriminator Generative Adversarial Network (GAN) framework. This Deep-learning-geometrical-driven solution generates designs very similar to traditional topology optimization’s outputs in a fraction of time. Moreover, it allows an easy shape fine-tuning by a simple increase/decrease of the input geometrical condition (here the total-bar-count), a task that a traditional topology optimization cannot achieve.