This paper presents a new method based on iterative improvement to generate the random microstructure of continuous fibre reinforced composite with high fibre volume fraction (up to 70%), overcoming the jamming limit of the traditional Random Sequential Adsorption(RSA) method. The novelty of the proposed method lies in an iterative improvement of the generated microstructure through separating initially randomly set positions of the fibres. Without complex heuristic steps, this algorithm can generate different fibre distributions with fast speed and great simplicity. The offset magnitude of the separation process is adjusted according to fibre volume fractions to speed up the generation. Statistical analysis is then performed on the generated fibre distribution. Five descriptors including Voronoi Polygon Area, Nearest Neighbour Distances, Nearest Neighbour Orientation, Ripley’s K Function and Pair Distribution Function are used to compare the proposed algorithm and the RSA with completely spatial random (CSR) distribution for both short and far distances. Fibre distributions generated by this new algorithm are proved to have good randomness. An FEA example is also presented to predict the effective elastic property of the carbon fibre epoxy composites, and a reasonable agreement with the experimental result is achieved. The proposed algorithm provides a useful tool to generate micromechanical models that can be used to predict and understand the mechanical behaviour of fibre reinforced composites.