In the study of fracture conductivity evolution under stress using direct numerically generated fracture models, a key issue is the neglect of the real fracture's heterogeneous microstructure. To address this, the Brazilian splitting method is used to create fractures in various types of rocks. A 3D optical topography scanner then captures the actual fracture morphology and aperture information. This data forms the basis for establishing a contact mechanics model and a single-phase seepage model, which are used to study the evolution of the fracture flow field and permeability under continuous stress loading. The study also evaluates the applicability of traditional empirical formulas to real fracture cases. The findings are significant: ① The 25 mm×50 mm real fracture in Brazilian splitting shows obvious heterogeneity in the microstructure of the original aperture and surface roughness, which is obviously different from the fracture that reaches the average scale directly generated by numerical method. ② In the process of stress loading, the fracture aperture, contact area and spatial correlation length show different evolution characteristics in the x direction and y direction due to the fracture heterogeneity, and the control mechanism of permeability change is different; ③ When the traditional empirical formula is used to fit the stress-sensitive permeability evolution of fracture, the deviation of the fitting degree increases with the increase of the heterogeneity of fracture samples. This study suggested that the traditional empirical formula has a good application basis in the study of the fracture reaching the averaging scale, but it is limited in the application of the fracture with strong heterogeneity or failure to meet the averaging scale.