To address the limitations of fixed step sizes and insufficient goal guidance in Informed-RRT*, this paper proposes AFLA-IRRT*, the path planning algorithm that combines the FLA-Net-based guided sampling network with the multi-factor adaptive step-size mechanism. FLA-Net employs farthest point sampling and multi-scale feature aggregation to predict path likelihoods within the ellipsoidal domain, enhancing informed sampling. During expansion, the algorithm dynamically adjusts step sizes based on obstacle proximity, goal direction, and sampling geometry, improving efficiency in complex environments. Experiments show that FLA-Net surpasses PointNet++ and CNN in loss, accuracy, and path prediction. AFLA-IRRT* further outperforms RRT*, Informed-RRT*, and Neural-IRRT* in path length, smoothness, and computation time, demonstrating superior adaptability and robustness.