To enable efficient prediction of transition heat flux fields under diverse freestream conditions, a generative transition heat flux prediction model based on variational autoencoder architecture was developed. The hypersonic cone configuration under different freestream conditions was selected as the research object, with numerical simulation method being employed to generate the transition heat flux dataset. A variational autoencoder model was constructed and was trained and validated on the transition heat flux dataset. The analysis of results demonstrates that the latent variables of the heat flux field can be effectively extracted by the variational autoencoder model, and the heat flux structure of the transition process induced by leeward-side streamwise vortices was accurately reconstructed. A fully connected neural network model was established to construct a nonlinear mapping relationship between the freestream conditions and the latent variables of the heat flux field. By connecting the fully connected neural network model with the decoder part of the variational autoencoder model, a hypersonic cone transition heat flux prediction model was developed. The prediction results indicate that this model effectively learns the characteristics of heat flux distribution under complex transition mechanisms, achieves high prediction accuracy for heat flux under various freestream conditions, with errors not exceeding 0.024.