To enable efficient prediction of transitional heat flux fields under diverse freestream conditions a generative prediction framework based on variational autoencoder architecture was developed. The hypersonic cone configuration was selected as the research object, with Computational Fluid Dynamics simulations being employed to generate the dataset encompassing multiple freestream parameters. The variational autoencoder architecture was systematically trained and validated, demonstrating its capability to extract low-dimensional latent space representations of complex heat flux fields. Particularly, the model was shown to achieve high-fidelity reconstruction of streamwise vortex induced thermal patterns within the leeward-side transition region. A multilayer perceptron was subsequently implemented to establish the nonlinear mapping between freestream parameters and the latent space representations. The integrated prediction model was formed through cascading the multilayer perceptron module with the variational autoencoder decoder. Results demonstrate the prediction accuracy of the heat flux under different free stream is quite perfect, prediction errors not exceeding 0.024 in normalized mean squared error across all test cases, while successfully capturing heat flux distribution features under complex transition mechanisms.