Driven by the need for more accurate decoding of perceptual experiences from brain recordings, this study addresses the limitations imposed by traditional onedimensional analysis of fMRI data. Our goal was to maintain the spatial information of fMRI data by employing a two-dimensional cortical surface-based analytical framework, aiming to enhance fMRI representation of neural responses. After converting volumetric fMRI scans into 2D surface data, we use Vision Transformers and Latent Diffusion Models to learn from fMRI and generate precise image. Our model achieved better representation abilities and generated clearer, more accurate high-resolution natural images from 2D fMRI inputs.