Abstract: To address the issues of jagged edges, aliasing artifacts, and detail loss in complex scene reconstruction with existing 3D Gaussian Splatting (3DGS) techniques, an adaptive frequency filtering method based on geometric features is proposed. This method dynamically optimizes the frequency response of Gaussian primitives to balance high-frequency detail preservation and low-frequency noise suppression. By analyzing the local geometric complexity, color gradients, and neighboring primitive density of the scene, a lightweight learnable mapping function was designed to dynamically associate multimodal features with the filtering strength parameters of Gaussian kernels, achieving frequency-adaptive 3D reconstruction. Specifically, we first extract three input features for each Gaussian primitive: the intensity of normal variation, color gradients in projected regions, and local neighborhood density. These features are fed into a Multilayer Perceptron (MLP) to predict isotropic smoothing coefficients, which are then integrated into the covariance matrix to adjust the filtering range. Additionally, a two-stage end-to-end optimization strategy is proposed to jointly optimize the attributes of Gaussian primitives and filtering parameters, combining RGB reconstruction loss, structural similarity loss, and regularization constraints to ensure training stability. To validate the effectiveness of our approach, comprehensive experiments are conducted on the Mip-NeRF360 dataset. The results demonstrate that the improved algorithm significantly outperforms existing methods in both reconstruction accuracy and visual quality.