This study serves to advance AI-based neural decoding, in which visual experiences are reconstructed from fMRI signals. The standing two-stage decoders utilize simple linear ridge regression for fMRI-to-latent mapping. To elaborate, brain activity is mapped onto a VAE latent space for structural features, then reconstructions are refined with diffusion models modulated by CLIP embeddings for semantic details. In this paper, the authors propose to replace the traditional ridge regression model with GRU-based non-linear mapping and optimize the latent space dimensionality in order to make the process more efficient. Through conducting experiments using the Natural Scenes Dataset, composed of 73000 images, and fMRI of eight subjects, it was found that the novel proposed architecture reduced complexity of the first stage by 85% while maintaining accuracy. Sensitivity analyses revealed that the first stage is crucial for structural integrity while the second stage served to contribute substantial semantic quality. Conclusively, this research not only paves the way for novel advances that improve efficiency of neural decoders, but highlights the ways in which such models correspond to the hierarchical processing of vision in the brain.