 The study presents a two-stage scene reconstruction framework called brain diffuser that uses latent diffusion models to generate high-complexity images from fmi signals, outperforming previous models on the natural scenes dataset benchmark and creating compelling ROI optimal scenes consistent with neuroscientific knowledge. This article was authored by Furkan Osalik and Rufin VanRullen.