In the field of geophysical oil and gas exploration, the velocity of seismic wave is regarded as one of the key parameters to distinguish underground media. Full Waveform Inversion (FWI) is a popular geophysical imaging technique in which a high-resolution subsurface velocity model is learned by first giving an initial velocity model and then minimizing the deviation between the simulated seismic record generated by the velocity model and the actual observed seismic record through an iterative process. The initialization of accurate velocity model is a prerequisite for the realization of full waveform inversion imaging, which is strongly dependent on the initial velocity model. When there is a big difference between the initial velocity model and the real model, the matching between the theoretical wave field and the observed wave field is easy to produce the phenomenon of period hopping. This paper proposes a full-waveform inversion model (FWI with initial velocity learnable by neural network, FWI_IVLNN)based on generative adversarial networks that can learn initial velocity parameters. It learns initial velocity parameters through neural networks and integrates stochastic velocity to generate velocity models. Based on generative adversarial idea, the velocity model is inversion, the simulated seismic record is generated by physics generator, the simulated record and the observation record are discriminated by discriminator, and the initial velocity model and network parameters are learned by unsupervised mode. The gradient returned by the network is transmitted to both the neural network and the physical generator, and the data-driven and model-driven velocity model inversion is realized with the help of the gradient. Quantitative and qualitative experiments show that FWI_IVLNN is less dependent on the initial velocity model, and the accuracy of velocity model inversion is higher.