Modeling neural responses under naturalistic visual stimulation is an important goal in computational neuroscience and brain-computer interface research. Progress in this area depends on neuroimaging datasets that combine repeated measurements, shared stimulus anchors, and sufficient stimulus diversity for evaluating encoding and decoding models. Here, we present the Natural Vision Dataset (NVD), a publicly available 5.0 T fMRI dataset designed for static natural-image viewing. The field strength is reported as part of the acquisition context, and the dataset was not designed to isolate field-strength effects or to compare 5.0 T performance with 3 T or 7 T acquisitions. Twenty healthy participants viewed a shared set of 1,268 natural images and 500 participant-specific images per participant, yielding 10,000 participant-specific images across the dataset. Each image was presented three times across separate sessions. This hybrid shared-unique stimulus design supports assessment of response reliability, cross-participant alignment, and model generalization beyond a fixed shared image set. Functional data were acquired at 1.8 mm isotropic resolution with a TR of 1 s and are organized according to the Brain Imaging Data Structure, with both volumetric and surface-based derivatives provided. Image-level response estimates were derived using GLMsingle toolbox. Data quality and benchmark utility were characterized using motion, vigilance, tSNR, noise-ceiling, and brain-to-CLIP decoding analyses. NVD provides a standardized resource for investigating human visual representations and evaluating computational models of natural-image processing.