Abstract : This work focuses on the linear elastic and thermal properties of real and virtual, computer-generated fibrous composites. A stochastic microstructure model is used to generate densely-assembled 3D systems of curved, non overlapping fibers with specific orientation distributions. This model is first optimized to approach the characteristics of a real fiber glass polymer by fitting geometrical and statistical parameters, such as fiber orientation, radius, length, and curvature. Second, random realizations of the stochastic models that depart from the characteristics of the fiber glass polymer are generated. The latter, which range from isotropic to transversely isotropic and to orthotropic materials, represent plausible virtual fibrous materials. Full-field numerical computations, undertaken by means of the Fourier-based (FFT) method, are used to estimate the local and effective mechanical and thermal responses of the fibrous composites. The anisotropy of the macroscopic responses as well as the size of the corresponding representative volume element (RVE) are examined numerically. It is found that the variance of the properties on a volume V scales as a powerlaw ∼1/V^α where α<1, an effect of long-range correlations in the microstructure. Finally, the overall behavior of the fiber composites are computed for varying fiber curvature and orientation distributions, and compared with available analytical bounds. We find that the fiber arrangement strongly influences the elastic and thermal responses, less so for the fiber curvature.