Abstract: Prestack seismic Amplitude Variation with Offset (AVO) inversion is an important technique for obtaining subsurface elastic parameters and supporting complex reservoir prediction. However, because of the band-limited nature of seismic data, noise interference, and multi-parameter coupling, the inversion process exhibits significant nonlinearity and ill-posedness. Conventional regularization methods, which typically rely on initial models and manual priors, incur high computational costs. Although supervised learning methods possess strong nonlinear mapping capabilities, they depend heavily on large-scale, high-quality, labeled data, which are difficult to obtain from actual well logs. To address this issue, this paper proposes a prestack seismic AVO inversion method based on implicit neural representation and semi-supervised fusion. This approach incorporates low-frequency prior information into the iterative optimization loss function, enhancing the constraints on low-frequency background information and improving the stability of the inversion process compared to conventional regularization methods. Additionally, by integrating a U-Net architecture with a semi-supervised learning strategy, the method embeds constraints from limited labeled data into the implicit neural representation process, thereby enhancing model representation capability and improving inversion accuracy while reducing reliance on labeled data. Finally, the proposed method and conventional methods are tested using two-dimensional synthetic data. The test results demonstrate that the proposed method improves the accuracy and resolution of elastic parameter inversion results compared with conventional methods, validating its effectiveness in prestack AVO inversion.