In computational multispectral metasurface design, conventional approaches typically rely on parameter sweeping and manual selection to identify array combinations with low spectral correlation, which is time-consuming and inefficient. We propose a global inverse design framework for computational multispectral metasurfaces that in-tegrates a conditional diffusion model with a genetic algorithm. In this framework, geometric parameters are en-coded into two-dimensional RGB images, enabling unified generation and representation of diverse structures, in-cluding cross-shaped, X-shaped, square, and C4-symmetric random patterns. The genetic algorithm is first employed to select spectral combinations with low correlation, which are then fed into the inverse design model to predict the corresponding geometries, thereby avoiding exhaustive search. The results demonstrate that the proposed method achieves a spectral reconstruction accuracy with a mean squared error (MSE) on the order of one thousandth. Overall, the proposed approach significantly improves design efficiency and provides a practical and scalable so-lution for compact and high-performance computational multispectral metasurfaces.