The integration of artificial intelligence (AI) in architectural design, especially for generating floor plans, can greatly streamline the process. However, most AI-generated plans focus on form and spatial layout, often neglecting crucial performance evaluations because they are presented as images without the necessary geometric and physical properties for effective simulation. To address this limitation, we propose a novel approach that combines diffusion models with generative adversarial networks (GANs) for generating and evaluating floor plans. We fine-tuned a Low-Rank Adaptation (LoRA) model for creating residential floor plans, while a GAN quickly predicts daylighting performance. Our results show that the diffusion model generates a more varied set of floor plans compared to the training set. The GAN accurately assesses daylighting performance, with deviations from the ground truth not exceeding 5%, achieving a mean squared error (MSE) of 4.2 and a structural similarity index (SSIM) of 0.98. Additionally, it operates 267 times faster than traditional methods. This approach equips architects with a reliable tool for efficient early-stage design decisions, enhancing AI-driven workflows.