Abstract: High-fidelity flow data are expensive to acquire, and the required sampling scale increases rapidly with the dimensionality of the parameter space, making full-domain high-fidelity simulations impractical for engineering applications. A bi-fidelity transfer learning framework based on an autoencoder (AE) was proposed to achieve cross-fidelity data fusion using only a limited number of high-fidelity samples in high-dimensional parameter spaces. First, abundant low-fidelity flow-field data were collected within the parameter space of interest to pretrain the AE, compressing parameterized flow fields into a compact latent space. Subsequently, rank-revealing QR decomposition was performed on the latent representation matrix to identify the most representative parameter points, at which high-fidelity simulation data were obtained. Finally, the encoder was frozen and the decoder was fine-tuned using a small number of high-fidelity flow fields to establish a cross-fidelity mapping from low-fidelity to high-fidelity data. The proposed framework was validated using a helicopter hover downwash flow-field dataset. By utilizing high-fidelity samples corresponding to only 17.4% of the total operating conditions, the reconstructed flow fields achieve an average relative error of 8.25% and an average error reduction rate of 45.84%. The proposed approach provides an effective solution for low-cost and high-accuracy modeling of flow fields in high-dimensional parameter spaces.