The deployment of autonomous robots in human-centric environments requires sophisticated scene understanding and adaptive visual processing capabilities. Current robotic perception systems rely predominantly on photorealistic 3D reconstruction, limiting their ability to adapt visual representations for task-specific reasoning or human-robot interaction contexts. While recent advances in 3D Gaussian Splatting have enabled real-time scene rendering for robotic applications, existing approaches for adaptive visual stylization remain computationally prohibitive for mobile robotic platforms. We introduce a principled approach to 3D style transfer that reconceptualizes feature decoding through geometric locality rather than volumetric processing. Our Enhanced Multi-Layer Perceptron architecture leverages the intrinsic manifold structure of 3D point distributions, performing adaptive feature aggregation via K-nearest neighbor graphs coupled with learnable transformations. By reconceptualizing feature decoding through K-nearest neighbor graphs rather than volumetric convolutions, our approach achieves the computational efficiency required for autonomous operation while preserving spatial coherence essential for robotic navigation and manipulation tasks. Experimental validation on diverse robotic scenarios demonstrates substantial efficiency improvements. Our methodology demonstrates robust generalization across diverse geometric complexities and artistic domains.